Research on vehicle detection algorithm based on improved YOLOV5

Given the large cost of the current vehicle detection implementation project, this paper reduces the cost of the project by modifying the Yolov5s backbone model to greatly reduce the number of parameters in the training process under the premise of ensuring accuracy. First, MobileNetV3 is used to replace the Yolov5s backbone network, which reduces the number of parameters from 14.4M to 3.1M and the mAP from 0.984 to 0.977. The accuracy of the model is only reduced by 0.007 with the greatly reduced number of parameters, and the model is greatly improved. The SE attention module of the inverted residual structure is replaced by the CBAM module, and the parameters and mAP are further optimized, thus reducing the parameter quantity to 2.4M and the mAP to 0.978, which makes the model better improved and can better meet the vehicle detection put into use.