*4.3. Single Shot MultiBox Detector Deep Learning Model Training Outcomes*

Three hundred twenty job site activity images, accounting for 80% of the data set, were selected as the training sample for the SS-based job site activity image recognition system proposed herein. In addition, 40 images, or 10% of the data set, were chosen as the test samples during the training. In the end, 61 images the model had not seen were used for recognition; thus, 461 images were collected and used. The visualization outcomes after recognition are presented in Table 5.

**Table 5.** Outcomes of single shot multibox detector image recognition model test.

Automated generation of EXCEL forms for the recognized results included object names, confidence level, pixel coordinates, and time record. The timestamp was based on the computer time when the form was generated, which could be used as the basis for specific management items (Table 5):


Construction activities at a job site vary widely. The machines subject to image recognition are excavators, loaders, dump trucks, cranes, and concrete mixer trucks, and the recognition accuracy is 69%, on average. The workers are wearing work clothing and reflective vests without a uniform standard, and they are at various locations within the job site performing various tasks, resulting in difficulties in recognition due to the bright side, dark side, and body position, and the recognition accuracy is 53%. The accuracy is 28% for the rebar. The reason for the low recognition accuracy could be that they are similar materials divided into two different classes; also, there are more than civil work activities at the job site; for example, there are plumbing and electrical tasks at a job site, and their materials, such as pipes and cables, may affect the recognition results, as shown in Table 6.

**mAP Recall (Threshold = 0.5) Precision (Threshold = 0.5) F1-Score (Threshold = 0.5)** Rebar 0.29 0.09 1.00 0.17 Worker 0.53 0.37 0.86 0.52

Machine 0.69 0.62 0.95 0.75

**Table 6.** Model performance indices.

This study uses automatic identification of construction site workers, material locations, and construction environment conditions of equipment. The resulting photos can identify more than two items simultaneously, providing site supervisors with active warnings of potential occupational safety hazards and increasing construction efficiency through image automation.
