*4.3. Experimental Results of Embedded Platform*

To further verify the processing capability of the improved model in mobile devices, the trained model is deployed to the Jetson Xavier NX embedded platform for verification. The processor is small in size, low in power consumption, and strong in computing performance. The performances of the YOLOv4 network model, the YOLOv4-tiny network model and the improved model in this paper are compared in terms of the objective evaluation indicators mAP and FPS respectively, as shown in Table 8. It can be concluded that for the Jetson Xavier NX embedded platform, the input image is 256 × 256 pixels, and the YOLOv4 network model can process 16 FPS due to its complex structure, which cannot meet the needs of mobile devices for real-time crack detection. The YOLOv4-tiny network model and the improved model in this paper can process 56 and 44 FPS, which can meet the needs of real-time detection. However, the mAP of the YOLOv4-tiny network model is 72.22%, and the recognition rate is low. Compared with the YOLOv4-tiny network model, the improved YOLOv4 network model has higher accuracy and faster processing speed, which meets the requirements of accurate real-time object detection.


