Research On Improved Lightweight Traffic Object Detection Model Based On YOLOv5

With the continuous improvement of China’s economic and social development level,modern transportation is also shifting from the development concept of focusing on scale and speed to quality and efficiency as the core.The wave of artificial intelligence technology has made great contributions to the construction of a modern comprehensive transportation system,and the development of emerging interdisciplinary disciplines,including intelligent transportation and unmanned driving,continues to facilitate people’s lives.As one of the key technologies,object detection needs to have the ability to serve intelligent transportation systems well and be widely used in various scenarios.However,traffic target detection models based on deep learning often have the characteristics of large number of parameters and computation,which is difficult to apply in some traffic scenarios with limited computing and storage resources.Therefore,it is of great significance to study a lighter traffic object detection model,and this paper will improve the lightweight object detection model based on deep learning,the main work content is as follows:(1)In view of the large number of parameters,high computational complexity and low running speed of object detection model based on deep learning,the research status of object detection in traffic scenes at home and abroad is analyzed,as well as the current compression method of neural network model.Through comparison,YOLOv5 s model with better speed and accuracy is selected as the benchmark model for target detection of lightweight traffic scenes in the subsequent research.Firstly,MobileNetv3 lightweight network was combined to replace the backbone network of the original model,and PASCAL VOC and BDD100 K data set were used for training and verification.In order to balance the effect of reducing the accuracy caused by replacing only lightweight network structure,a fusion of Focus module and MobileNetv3 network layer is proposed,and the network accuracy is restored by replacing CIOU loss with EIOU loss.The experimental results show that the parameters of the improved network model are significantly reduced compared with the original YOLOv5 s model and the calculation amount,and the accuracy of the model is maintained while the lightweight is improved.(2)The Gamma parameter size of the BN layer is used to judge the importance of the channel,and the network model after the sparsity training is structured pruning operation,which solves the problem that the improved network model still has certain redundant parameters.On the premise of not obviously reducing the accuracy of the network model,the parameter number and calculation amount of the improved model are further reduced.The experimental results show that,compared with the original YOLOv5 s model,the pruned network model reduces the number of parameters by68.9%,the amount of computation by 25.6%,the detection frame rate by 46.4%,and the real-time detection of traffic targets is improved when the average accuracy mAP decreases by 1.3%.(3)The improved lightweight traffic object detection model was tested on the NVIDIA Jetson Nano embedded platform by using the TensorRT neural network inference acceleration framework,aiming at the problem that the target detection model is limited in the deployment of low-power platforms.The experimental results show that compared with the original YOLOv5 s model,the improved lightweight model is accelerated by the Tensorrt reasoning,and the final average frame rate reaches 24.1fps,which basically meets the real-time requirements.