Target Detection Algorithm Research For Automatic Driving Oriented Traffic Scenarios

In the field of computer vision research,automated driving technology is one of the current mainstream research directions.The traffic scene target detection algorithm is a key concern in the field of artificial intelligence.In practice,traffic scene targets usually refer to dynamic targets such as pedestrians and vehicles,and vehicles can be classified as bicycles,cars,buses and other types with different appearances.Along with the development of artificial intelligence,deep learning target detection algorithms have demonstrated their irreplaceable excellence in today’s era and have been practically applied in many real-world products,including the autonomous driving field,which relies on accurate target detection algorithms to be implemented.However,it has been shown through numerous experiments that the model accuracy of a network is usually related to the size of the model.In order to improve the detection accuracy of the network model,researchers often choose to deepen the level of the neural network and enhance the complexity of the network structure to make it have stronger feature extraction capabilities.This leads to disadvantages such as the large training parameters required for the network and the large burden on the hardware configuration for a long running time.Traditional target detection algorithms usually rely on high-end hardware devices.However,in the field of autonomous driving,hardware chips need to be deployed on the mobile side for edge computing.Compared with GPUs and TPUs deployed on computers,the chips that can be equipped on the mobile side often do not perform as well as expected.Complex neural network structure,heavy training parameter calculation are huge challenges for the chip.In summary,in this paper,the dynamic target detection represented by vehicles and pedestrians is targeted for improvement research,and the main work is as follows:An improved YOLO model based on the lightweight concept is proposed.The current mainstream network development usually focuses on accuracy as the primary goal,and the network model becomes more and more complex,the number of parameters becomes larger and larger,the overall memory consumption becomes more and more,and the time required for detection becomes longer and longer.In this paper,the YOLO(You Only Look Once)v5 target detection algorithm is used as the basis,and its backbone network is redesigned with various lightweight networks such as Mobile Net series,Shuffle Net,Rep VGG,etc.,and the excellent network model is summarized by comparative analysis.The method in this paper is experimentally validated,and the network parameters are significantly reduced and the training speed is improved.(1)Using Mobile One idea to perform secondary lightweighting of the improved network in this paper,and using Conv Next structure to improve the neck part of YOLOv5 to enhance the feature enhancement capability of the network,and using SPD convolution to improve the extraction and enhancement efficiency of the lightweight network for small target fine-grained information,and finally verifying the balance between lightweighting and recognition accuracy by ablation experiments.