Design of a Low-Cost AI System for the Modernization of Conventional Cars

Abstract

Artificial intelligence techniques are beginning to be implemented in most areas. In the particular case of automobiles, new cars include integrated applications, such as cameras in different configurations, including in the rear of the car to provide assistance while reversing, as well as front and side cameras; these applications also include different configurations of sensors that provide information to the driver, such as objects approaching from different directions, such as from the front and sides. In this paper, we propose a practical and low-cost methodology to provide solutions using artificial intelligence techniques, as is the purpose of YOLO architecture, version 3, using hardware based on Nvidia’s Jetson TK1 architecture, and configurations in conventional cars. The results that we present demonstrate that these technologies can be applied in conventional cars, working with independent power to avoid causing problems in these cars, and we evaluate their application in the detection of people and cars in different situations, which allows information to be provided to the driver while performing maneuvers. The methodology that we provide can be replicated and scaled according to needs.

1. Introduction

Currently, with the development of technology, we are witnessing advances in process automation, thanks to the growth of artificial intelligence. The automotive sector is being influenced by these advancements, as seen in the design of cars that operate independently of combustion engines and the use of automation techniques based on cameras and sensors that provide drivers with aids to improve their driving experience. In reviewing the state of the art, we found many technological solutions related to incorporation into cars.

In a review of the literature, we found several applications that make use of artificial intelligence techniques, with the most commonly used being YOLO architecture, which allows the detection of objects through video analysis. We present related work where a comparative study of six deep learning models for the detection of potholes in pavements was conducted, including variants of YOLO such as YOLOv3, YOLOv4, and YOLOv5. It was found that YOLOv4 achieved the best performance, with a mAP of 83.2%, highlighting its ability to detect small potholes more accurately than YOLOv5 models. It was suggested to improve label quality, increase training data, and adjust hyperparameters to potentially further improve pothole detection accuracy in future research [1].

The STC-YOLO model was used for the detection of small traffic signals in complex environments, employing the CIoU loss function and NWD location loss calculation to improve accuracy. The Swin Transformer block was implemented to enhance computational efficiency, and results were compared with other models on the TT100K dataset, demonstrating advantages in accuracy and real-time performance. Additionally, it is highlighted that the STC-YOLO model achieved significant improvements in small traffic signal detection, with an increase in mAP from 79.6% to 86.4% when adding the MPANet module. Combinations of modules such as C4STB, NWD, and K-means++ also showed improvements in model performance, demonstrating that each implemented module provides an enhancement in the performance of the object detection network [2].

The development of an architecture and implementation of a system to count people and bicycles in real time using YOLO on Jetson Nano is presented. The system was divided into video input, object detection, and tracking modules, with a focus on optimization for low-power devices. Different object detection models were compared, highlighting the superior performance of models optimized by TensorRT. In addition, the possibility of implementing and testing new models and algorithms in the future is mentioned. The work was developed in collaboration with the Polytechnic Institute of Coimbra and focuses on multiple-object tracking accuracy, with metrics such as MOTP and MT. It highlights the efficiency of the system in maintaining real-time execution with reduced memory usage when performing object counting tests. Reference is made to previous work in the field of object detection and tracking, and the implementation of the project from scratch is highlighted, with the exception of the use of the YOLOv5 model and the object tracker [3].

An M-YOLO algorithm was designed for traffic signal detection, which combines modules such as SPPNet, CPSNet, and FOCUS to improve the accuracy and speed of small target detection. Compared with YOLOv3, the proposed algorithm increased the precision value by 8.6%, recall value by 0.5%, and mAP value by 3.3% on the CCTSDB dataset. These results demonstrate a significant improvement in traffic signal detection, which is crucial for road safety and intelligent transportation systems. In tests performed on the CCTSDB dataset, it was observed that the proposed algorithm was significantly improved, showing an increase in precision, recall, and mAP compared to YOLOv3. Traffic signal detection in normal and challenging environments is critical to ensuring road safety, and the use of specialized small target detection modules has proven to be effective in improving traffic signal detection in complex situations [4].

An analysis intending to compare and assess the performance of four YOLO models on three edge intelligence devices, including Jetson Nano, Jetson Xavier NX, and Raspberry Pi 4B with NCS2, was conducted. The inference performance of YOLOv3 and YOLOv4 on these devices was evaluated, yielding empirical recommendations for deploying AI applications on edge devices based on the performance of SBCs. This analysis provides valuable information for the development of intelligent applications in edge environments, highlighting the importance of proper hardware selection in optimizing the performance of AI models [5].

A novel, lightweight approach to real-time traffic signal detection, crucial for road safety and urban traffic efficiency, was presented. The combination of YOLO and MobileNet was highlighted to optimize feature extraction, reduce computational burden, and improve real-time performance. In addition, a hierarchical feature interaction structure was proposed to facilitate multi-scale feature fusion and reduce dependency on the hardware environment. In summary, the study addresses the challenges of traffic signal detection under varying conditions and proposes an efficient and lightweight integration framework based on YOLOv4 and MobileNet to improve accuracy and real-time performance with significant parameter reduction and an innovative feature interaction structure [6].

The implementation of an IoT-based autonomous parking system to optimize the use of parking space and reduce traffic congestion in smart cities was discussed. It highlights the importance of using machine learning and transfer learning techniques to adapt the system to different urban environments, promoting scalability and efficiency. The research focuses on improving the user experience by automating the detection of empty and occupied parking spaces, with the potential to benefit both local residents and the overall urban community [7].

Neural network methodologies such as SSD and YOLO were utilized for accurate vehicle detection and classification in traffic images. A comparison of algorithms is highlighted, where YOLOv7 showed an accuracy higher than 99% in low-light conditions, followed by YOLOv6. The importance of these structures for vehicle detection efficiency is emphasized, with potential applications in surveillance and intelligent transportation systems. In addition, the availability of challenging datasets, such as Shah’s Dataset III, which includes complicated traffic flow types, is mentioned. A summary table of the selected datasets in terms of pixel resolution and recording time is provided, along with detailed explanations on the choice of these datasets. This highlights the relevance of the presented methodologies for vehicle detection and classification in varied traffic environments [8].

The DAMP-YOLO model enhances the YOLOv8 network by incorporating modules such as deformable CSP bottleneck, aggregated attention, and data augmentation for meter reading recognition. The efficiency of the model in complex environments and the significant reduction in the model size are highlighted. In addition, ablation experiments were performed, demonstrating the positive contribution of each proposed module on the overall model performance. In summary, DAMP-YOLO excels in detecting objects with similar shapes and structures, thanks to its innovative modules and network pruning techniques. Experimental results showed a substantial improvement in precision, recall, and mAP compared to other state-of-the-art methods, positioning it as a promising solution for real-time meter reading recognition, especially in resource-constrained devices [9].

A visual system for detecting children and pets in closed vehicles was developed using data augmentation algorithms and evaluating object detection models such as YOLO6 v1, YOLO6 v2, YOLO7, and NanoDet, with YOLO6 v1 demonstrating the highest accuracy. The proposed system sends notifications to parents when it detects children or pets without adults in the vehicle, thus addressing the prevention of hyperthermia and suffocation incidents. The detection models were analyzed in terms of false positives and false negatives, highlighting the effectiveness of YOLO6 v1 and the importance of accurate detection for the safety of vehicle occupants, especially in critical situations such as the presence of unsupervised children or pets [10].

Significant improvements in the YOLOv5s model for real-time object detection are introduced, incorporating BiFPN to fuse features of different scales and optimizing the PBE-YOLO model to balance accuracy and computational efficiency. The GIoU loss function was replaced with EIoU to improve convergence speed, resulting in improved performance in railroad turnout identification. Ablation experiments validated the enhancement modules in the lightweight backbone network, the feature fusion method, and the regression loss function, demonstrating that YOLOv5s_PP-LCNet outperforms other lightweight networks in accuracy and efficiency. In summary, the improvements implemented in YOLOv5s, such as the introduction of BiFPN and the optimization of the PBE-YOLO model, proved to be effective for railway turnout detection, achieving a balance between accuracy and computational efficiency. Experimental results showed favorable performance in object detection, highlighting the superiority of YOLOv5s_PP-LCNet over other lightweight architectures in terms of accuracy and efficiency [11].

An improved image fog removal model based on a conditional generative adversarial network (cGAN) is proposed. This model includes two modifications to improve the clarity of images when removing fog. Fog removal datasets were compared to validate the effectiveness of the proposed model, showing that it outperforms other deep learning-based methods. In addition, the enhanced images were used for object detection with the YOLO object detector, showing significant improvements in detection accuracy, especially in simulated weather conditions such as rain and fog [12].

A comparison of object detection architectures in occupancy grid maps using backbones inspired by YOLOv2, YOLOv3, and PIXOR was conducted. Key architectural differences between Darknet-19, Darknet-53, and CSPDarknet-53 are highlighted, and precision–recall curves for different detectors are presented. In addition, average precision scores and real-time capabilities of the proposed detectors are evaluated, demonstrating the feasibility of real-time vehicle detection in occupancy maps. In summary, the study focuses on object detection in occupancy grid maps, comparing different architectures and evaluating their performance in terms of accuracy and real-time capability. The proposed detector architectures and their differences with respect to the used backbones are discussed in detail, and quantitative results supporting the effectiveness of these approaches for real-time vehicle detection are presented [13].

A lateral control method for intelligent vehicles combining feedforward and predictive LQR algorithms is proposed. The system includes an LQR controller, feedforward controller, and predictive controller based on a vehicle-dynamics tracking error model. An environmental awareness system was implemented using the YOLOv3 algorithm to collect environmental information and adjust the vehicle tracking trajectory. Simulation and hardware tests confirmed the effectiveness of the proposed control for real-time path tracking, improving the accuracy and effectiveness of vehicle control. The study focuses on the lateral control of the vehicle, leaving aside longitudinal control, which will be considered in future research to establish a more complete tracking control system [14].

The implementation of a deep learning-based model to detect road anomalies and prevent traffic accidents is discussed. Using YOLO (You Only Look Once), vehicle collision types were classified, focusing on head-on collisions. A service that notifies in real time about accidents and obstacles on the road is proposed, using CCTV and AI to improve road safety and the autonomous vehicle sector. Simulations and experiments were conducted to verify the effectiveness of the service in cities such as Daegu, Busan, and Gwangju [15].

In the development of lane and object detection algorithms implemented in GTA5 and tested in DonkeyCar for real-time feasibility, a simplified implementation of digital twin technology for autonomous vehicles is proposed. The efficiency and low memory consumption of the algorithms are highlighted. The importance of validating the algorithms by HILS simulation and co-simulation for complex systems is discussed, demonstrating their applicability in real and virtual environments [16].

A proposed method for detecting vehicles in tunnels aims to improve road safety. A pre-trained VCG-16 model and the YOLO v2 detector were used for learning, optimizing detection with noise reduction and soft illumination in tunnel images. The comparative results show the effectiveness of the proposed method in detecting vehicles in this specific environment, highlighting the importance of this technology in preventing traffic accidents in tunnels and improving road safety management in these critical areas [17].

An improvement in vehicle detection in unmanned aerial vehicle (UAV) imagery using the YOLOv3 algorithm is presented. A vehicle dataset was constructed, and a YOLOv3 vehicle detection framework is proposed that excels in handling the small size and density of targets in UAV imagery. By using K-means++ and Soft-NMS, an increase in AP by 5.48% was achieved while keeping the processing speed high. Experimental results demonstrated a better detection capability than other methods such as Faster R-CNN. In summary, the implementation of enhancements such as K-means++ and Soft-NMS in YOLOv3 resulted in a significant improvement in the accuracy and detection rate of vehicles in UAV imagery. Validation on different datasets confirmed the effectiveness of this approach, highlighting the enhanced network’s ability to identify small and dense targets with increased accuracy and processing speed [18].

An improved vehicle detection algorithm focused on YOLOv3, proposing a GIoU loss function to optimize its performance. Challenges in vehicle detection are discussed, and algorithms such as Faster R-CNN and YOLO are compared. The importance of balancing datasets and improving the generalization capability of models through methods such as mixup is highlighted. In addition, a distributed system based on smart streetlights for vehicle detection is presented, demonstrating the practical application of the proposed improvements. Data collection from a specific set of vehicle images is mentioned, and the effectiveness of the implemented enhancements in terms of accuracy and model convergence speed is analyzed. The combination of techniques such as GIoU and IoU is highlighted as key to improving vehicle detection in traffic environments [19].

The application of deep convolutional neural networks in traffic flow monitoring is discussed, highlighting the optimization of a YOLO algorithm to achieve results close to the actual number of vehicles. It is proposed to improve the YOLO-UA model by replacing its network structure, using focal loss to address category imbalances, and applying batch normalization during training to increase its adaptability to various weather conditions. In conclusion, the aim was to improve the accuracy and robustness of traffic flow monitoring in different scenarios and weather conditions by optimizing an improved YOLO model, achieving more accurate and reliable results in vehicle detection [20].

An integrated approach for motorcycle detection and distance estimation using a single camera is presented, improving the MD-TinyYOLOv4 algorithm and evaluating Monodepth2 models for depth estimation. An average accuracy of 0.28 m was achieved in motorcycle distance estimation at 46 fps, addressing previous limitations and offering a promising solution for improving motorcycle safety in autonomous driving and intelligent navigation applications [21].

Several works are related to the use of video-based classifiers to locate people, cars and trucks in various applications, such as optimization analysis in parking management to know how many cars are in the parking lot and available spaces [22]; some works are dedicated to detecting motorcycles through the use of new technologies such as video analysis [23]; and we also find works related to the analysis of traffic in strategic locations to interact with traffic light systems to optimize traffic control; and solutions based on improving the performance of the systems by giving autonomy to cars to operate without the presence of drivers [24].

In this work, we implemented YOLO architecture, version 3, on embedded hardware to be installed in conventional cars in order to apply object detection through online video acquisition. This can provide assistance depending on the location of the cameras: tests were performed by installing cameras in the front and in the side-view mirror, which can detect people in different situations, as well as various types of cars. We describe the processes for implementation, indicating the required hardware and software.

2. Materials and Methods

The materials and methods were designed to implement a low-cost solution based on low-power embedded hardware to run artificial intelligence applications; in our case, these applications were based on YOLO architecture, performing the recognition of people in their different manifestations and activities, as well as all types of cars. The solution we present can be implemented in any type of conventional car, allowing all cars to exploit the techniques provided by artificial intelligence.

In Figure 1, we present details of the proposed methodology, which is based on an analysis of the problem, a description of the proposal, and the uses and applications of low-cost technology. This methodology aims to enable upgrades to conventional car units so they canexploit the pattern detection techniques of artificial intelligence.

2.1. Analysis of the Problematic Situation

Currently, the new cars that are available have technology installed that helps and improves the driver’s experience, based on the use of different sensors and video cameras that provide drivers with information when driving and parking, among other actions. These aids are not found in conventional cars. Maintaining the logic that technology can help with various activities performed by people, with emphasis on the process of driving, it is important to implement these aids in conventional cars.

2.2. Description of the Proposal

The proposal concerns the use of low-cost and low-power-consumption embedded hardware, with which one can run pattern recognition programs in real time. We provide below a description of the hardware and software needed to exploit YOLO architecture, version 3, in connection with a conventional camera that can be installed in strategic areas of the car; in our case, we tested the proposal by installing a camera in the side-view mirror of a conventional car.

In Figure 2, we present the architecture of the solution, which has as its central part Nvidia Jetson TK1 hardware. We installed a common webcam, which was responsible for recording the movement of the car, complemented with the installation of a 12-inch touch screen, to allow the driver to independently interact with the device. To ensure that the solution could be used without the need for the electrical system of the car, we recommend the use of external power through an external power supply system using rechargeable batteries. The following figure details the architecture.

2.2.1. Hardware

For the replication of the proposal, we had to ensure a low-cost hardware architecture and low power consumption, as well as simple use and installation, allowing the execution of high-performance applications that run in real time. In our particular case, for the execution of YOLO architecture, version 3, we needed to perform pattern detection in real time by recording video obtained from a camera. We used embedded hardware from Nvidia, which had low power consumption and sufficient computational capacity to run the YOLO application. The model chosen was the TK1 version, which is one of the first versions of embedded hardware commonly used for artificial intelligence applications. It contains in its architecture a 4-core Cortex-A15 CPU and a Kepler architecture GPU with 192 Cuda Cores, 2 Gb of RAM DDR3 and a storage unit of 16 GB eMMC, and it has USB 3.0 connectivity and a HDMI connector. All of these components are included in the same hardware, so the model does not require additional devices to function, only a 12 Volt power supply.

Figure 3 shows the hardware required to work with Jetson Tegra TK1. As can be seen, it has a reduced size of 5” × 5” (127 mm × 127 mm), which can be installed anywhere in the car. Tt has all the necessary components to work, requiring just a screen, keyboard and mouse to be connected.

Central to the proposal is the easy use and management of the device. We resorted to the use of a 7” × 5” touch screen, which allows coupling with the Jetson Tk1 card. The screen was easy to connect using HDMI connectors for video transmission and USB for power. The display worked integrally with the board; in our case, we required access to the operating system and the execution of applications.

Figure 4 shows the screen in use, where one can see the operating system being accessed. Using touch, one could interact with the operating system and with the installed applications.

The successful application of the proposal consists of being able to analyze the video captured in the side-view mirror of the car, from where images of people and cars are recorded, while performing maneuvers during which the side-view mirror is used, as is the case when reversing. The screen should show the video that the camera is capturing with the recognition of people and cars.

Figure 5 shows the configuration of the camera that was installed in the side-view mirror of the car, for demonstration purposes. The camera was a low-cost conventional camera that could be installed with different operating systems—in our case, Ubuntu version 14. The operating system detects the camera automatically, without the need to use any particular driver.

The camera was aligned with the side-view mirror of the car in order to obtain the same image that is present in the mirror. If the driver moved the side-view mirror, the camera performed the same movement, so the camera registers an image identical to that in the side-view mirror.

Figure 6 shows the camera located in the side-view mirror, from a rear perspective, identifying that the camera was aligned with the side-view mirror of the car.

Figure 7 shows the installation of the solution inside the car for demonstration purposes. The solution was located on top of the gearshift lever so that the driver could view the image when the car was reversing. As can be seen, the card and the screen were integrated into a single component, which would allow its implementation in any type of car.

2.2.2. Software

The software component is a central component of the proposal, allowing the exploitation of hardware resources in different solutions. In our case, we required a video display in real time using a camera and the simultaneous performance of a process of pattern detection, mainly of people and cars. Regarding the technical characteristics of the Jetson TK1 card, this was pre-installed with Ubuntu version 14, which supports Cuda 6.0. OpenGL 4.4 and OpenCVtegra, which facilitated the installation of the following packages for the execution of the YOLO V3 architecture.

The application developed was executed with the Python programming language, version 3, together with the Pytorch library, as well as the following libraries, which allowed us to compile the program without problems to create our development environment.

• opencv-python=4.2.0.34;
• Pillow=7.1.2;
• pyparsing=2.4.7;
• PyQt5=5.15.0;
• python-dateutil=2.8.1;
• scipy=1.4.1;
• tensorboard=1.13.1;
• tensorflow=1.13.2;
• torch=1.5.0;
• torchvision=0.6.0.

With the packages installed, we proceeded to implement our solution based on YOLO V3 architecture, with images of people and cars, so we could train our own architecture. This procedure required a large number of images of people in their different manifestations and of different cars.

Below, we present the pseudocode of the implementation, where we describe the main procedures to be performed in order to have our YOLO V3 architecture trained to detect people and cars.

• Create the dataset of images and labels.
• Generate the YOLO V3 architecture.
• Configure the location of the dataset in the YOLO V3 architecture.
• Perform the training of the YOLO V3 architecture.
• Run online video object detection.

For the creation of the dataset, it must be taken into account that the model is VOC, which is required by YOLO V3 architecture. The data set had to contain as many images as possible of both people and cars, and each image had to contain its corresponding label.

One of the desired characteristics of the solution is the robustness of classification, for which we designed our own database of images of people and cars. This procedure began with the recording of videos using the same camera to be used in the solution. The recording was performed in the streets of the city of Lima, which, most of the year, presents a cold climate with abundant fog due to its proximity to the sea. Also, traffic is high, and there is a lack of safety awareness by pedestrians. Due to all these circumstances, Lima has a high rate of traffic accidents; therefore, the recorded videos present real cases of people and automobiles.

After the recording process, we proceed to decompose the video into images; then, we labeled the people and cars that were in the images, for which we resorted to the application “labelImg”, which, as a result, presented each image with its respective label in the format supported by YOLO to be able to train it.

For the generation of the YOLO V3 architecture, this can be programmed with Python and Pytorch, or a basic model can be taken from different platforms such as github, where you can find several models of this architecture, such as basic models and already trained models, which helped us in our implementation.

Once in possession of the architecture of YOLO V3, the location of the dataset must be configured, as well as the location of the images and labels. This configuration is important because YOLO needs to know where the images and labels are to be able to train itself.

Having configured the architecture of YOLO V3, we performed training to prepare our YOLO V3 architecture. We had to take into account that for this process, we could train the model with the Jetson TK1 card itself due to its computational capacity. This process can take time depending on the number of images that are in the dataset. Another option was to perform the training on a computer with greater computational capacity and then export the weights with the model already trained, thus optimizing the training process.

Finally, once our object detector based on YOLO V3 was trained, we proceeded to perform detection on video captured in real time, which was obtained directly from the camera that was located in the side-view mirror, which performs registration in real time. The detection has a parameter or detection threshold that can be configured in the application; this threshold is the level of detection that you want to give the detection. For example, you can select an 85% level of certainty: this means that when you have a detection level of 85% certainty that it is a person or car, a box is drawn on the person or car detected. Depending on the configuration, you can select security levels above 85%; for our case, with a level of 85%, we could classify without problems people in critical situations that were in blind spots in the projection of the side-view mirror, according to the images presented in the results section.

2.3. Uses and Applications

The proposal presented consists of being able to register and detect people and cars in a video recording from a camera located in the side-view mirror of a car. In this way, information regarding nearby people or cars can be provided when driving, or, more critically, when the driver is reversing. We now present several cases in which the proposed solution provides very useful and important information.

3. Results

The results we present are related to providing real-time detection in critical situations, where the driver in most cases cannot see people when driving or when reversing. These situations are the subject of the most common accidents, caused by a lack of aids that indicate the presence of people and cars.

Figure 8 shows the detection of people and cars. In this very particular case, there is a person working in a garden, whose presence is hidden by a nearby tree. This tree may keep the driver from seeing the person. If the car is required to move back, our solution is very important for this case, as it provides help because detection is performed in real time; meters further back, the solution also allows the detection of a car, even if the image is not very sharp, so the architecture is robust for these two classes.

In Figure 9, you can see the detection of people even when they are in contact with other objects even when they are riding bicycles, which allows verifying the functionality of the proposal.

Another of the most common accidents caused by collisions is when a car needs to back up and the driver cannot see nearby cyclists. With the proposal presented, it is possible to detect people who are in the coverage area of the camera, even when they are riding a bicycle, which demonstrates the high level of performance of the model presented; meters behind, a car is detected even though it is far away.

In Figure 10, you can see the detection of several people and cars at the same time, the people are performing different activities, which allows validating the functionality of the proposal.

Another cause of accidents is when, for different reasons, drivers cannot see road maintenance workers when performing maneuvers. The proposed solution can detect these people; this information will be very helpful to drivers when performing maneuvers. It also detects people who are on sidewalks near roads.

In Figure 11, you can see the detection of people in the normal action when walking on the sidewalk, the functionality of the proposal can be exploited by installing the camera in different positions.

The model also detects people on sidewalks, so when they need to perform maneuvers, especially when backing up, they can see them with the detector, which allows them to have an aid when driving.

In Figure 12, we can visualize that it also performs truck detection, which increases the performance level of the proposed model. Similarly, it detects cars that are hidden by other cars, thus further proving to be a robust solution.

Figure 13 shows one of the main objectives of the solution presented, which is the blind spots presented by conventional side-view mirrors, which depends on the location of the mirror and the driver’s ability to use the image presented in the side-view mirrors.

In images 1 and 2, the presence of automobiles in the mirror is evident due to the proximity of the automobile, in contrast to images 3 and 4, in which the cars are far away, which causes the famous blind spots of the side-view mirror, which presents a very high risk of possible accident.

The image is complemented by the fact that depending on the location of the side-view mirror, the cars can be seen in red and green, which will depend a lot on the driver’s ability to maneuver the car.

The solution presented, by using a camera that has greater coverage and thanks to artificial intelligence algorithms, can detect the presence of cars automatically and provide this information to the driver on a screen, allowing a greater spectrum of visualization, thus eliminating the risks associated with blind spots in side-view mirrors. Our proposed system can be applied and replicated in conventional cars, thus exploiting the techniques provided by artificial intelligence.

The tests performed are described with the configuration of the camera in the side-view mirror of the car, which presented classifications of cars and people at an average distance of 20 m. This distance can be increased by improving the resolution of the camera, which would increase the system’s ability to discriminate and classify. Due to the fact that the camera ran continuously, the system is considered to have performed classification in real time.

The implementation of the solution requires an investment in the Jetson TK1 card, which can be obtained for USD 190 on average. The battery system and the touch screen cost an additional USD 60, thus totaling USD 250. The Jetson TK1 card hardware allows the integration of various sensors, conceiving the implementation of solutions based on artificial intelligence.

4. Discussions

The following discussions are based on applications that use systems built on artificial intelligence, such as YOLO architectures and convolutional networks, which have a direct relationship with the proposed work. We start with the use of the YOLO architecture in parking management, where it is used to recognize cars and analyze available spaces, thus improving parking management. This allows for better control of the locations of cars and enhances the management of parking lots, for which cameras are located in strategic locations to have greater coverage of parking areas [9].

There are solutions based on convolutional networks for traffic detection in strategic areas of highways, where cameras are located along the center line of the road to detect cars and interact with traffic light systems to improve traffic flow [10]. In these cases, we use video classification of cars with live video recording from strategic locations. In the case of placing cameras in the car itself, we found systems that analyze the front view that the car faces when moving, which are trained to improve solutions that seek to provide autonomy to cars. In this particular case, YOLO architecture version 4 was trained with the MS COCO database [24].

In our work, we present an architecture based on YOLO version 3 to provide conventional cars that do not have integrated cameras with a system for classifying people and cars using video recorded by cameras located in the side-view mirror. A camera was installed in the side-view mirror to record video, providing the driver with a tool when maneuvering in reverse and aiming to eliminate blind spots. The intention of the work was to exploit the technology available for use in conventional cars. To train the YOLO architecture, we implemented our own database with people we found on the streets and various cars. The procedure began with recording using the same camera as the one to be used in the solution, followed by processing to extract images of people and cars in order to train with images that can be encountered when using the solution.

5. Conclusions

The conclusions we reached at the end of this research relate to providing low-cost solutions that have important applications in people’s lives. This time, we presented a low-cost system that detects people and cars using video recorded by cameras located in side-view mirrors, in very complex situations when driving or reversing. The results were presented in cases in which it is very difficult to observe people—for example, when they are in gardens performing maintenance, riding bicycles, or performing public cleaning work—as well as automobiles of various types, such as cars and trucks, in different views like from the front and side. This shows that the solution is practical in its implementation and useful for providing technological alternatives based on exploiting techniques provided by artificial intelligence, with emphasis on their application in conventional cars that do not have these new technologies incorporated.

From the point of view of the required hardware, we recommend the use of the Nvidia Jetson TK1 embedded board, which comes with exclusive technical features to run applications based on artificial intelligence. In the proposal, we recommend the use of a touch screen to allow the operation of the solution without other devices such as a keyboard or mouse. In Figure 13, we present the final configuration, focused on allowing the use of a 12-volt battery bank, independent of the car’s power line, thus not influencing the car’s performance. The low power consumption of the solution allows for optimized use, working with battery replacement and enabling continuous and safe operation.

In Figure 14, we present the final hardware configuration, where we can see the HDMI and USB connections for the display, as well as the battery container on the sides, where batteries can be installed for independent work.

Finally, we indicate that the solution can be replicated and scaled depending on the user’s uses and needs, as long as the hardware to be used is compatible with versions of Ubuntu—in our case, version 14.