Optimization of the Pedestrian and Vehicle Detection Model based on Cloud-Edge Collaboration

With the increase of delay-sensitive applications, it is a trend to deploy deep neural networks on the edge. However, to improve the accuracy of pedestrian and vehicle detection models, various types of detection models such as YOLOv4 have complex structures, resulting in a large number of computing resources required for model training, which in turn makes it challenging to deploy pedestrian and vehicle detection models on resource-constrained edge devices. A common method is channel pruning. On the one hand, the configuration requirements for the running environment are too high to be done by the edge alone. To solve this problem, a pedestrian and vehicle detection framework based on cloud-edge collaboration is proposed to facilitate the deployment of pedestrian and vehicle detection model at the edge. On the other hand, the pruning proportion needs to be set manually and the evaluation criteria are mostly one-sided, which leads to easy to delete channels by mistake. Therefore, an automatic pruning method combining the convolution layer and BN layer is proposed to compress the model size. Experimental results show that the proposed framework enables the pruned model to be deployed on Jetson TX2, and the FPS is improved by 6.72 times. Channel pruning reduces the volume and parameter number to 96.77%, and the computational amount to 81.37%.