Research On Vehicle-pedestrian Visual Object Detection Algorithm For Intelligent Road Side Edge Computing

Visual object detection is one of the key technologies of roadside perception in cooperative vehicle infrastructure systems,and it is also a continuous research hotspot in academia.Roadside perception has high requirements on the accuracy and real-time performance of visual object detection algorithms.However,the existing lightweight object detection algorithms have relatively low accuracy.Meanwhile,the computational complexity of general object detection algorithms is relatively large,and it is difficult to deploy on roadside edge computing devices to meet the real-time requirements of applications.It is necessary to use model compression technology to reduce the computational complexity of the object detection algorithms.In addition,the roadside scenes also have high complexity,and the object detection algorithms need to be adapted to the scenes for better detection performance.According to the performance requirements of roadside perception for visual object detection algorithms,this thesis studies and designs a vehicle-pedestrian visual object detection algorithm for roadside edge computing.The main tasks are as follows:1.This thesis analyzes the current mainstream framework of visual object detection algorithms and model compression algorithms,and summarizes algorithms including YOLOX,Res Rep channel pruning and knowledge distillation based on attention mechanism.It also summarizes the characteristics of roadside scenes and the performance requirements of roadside perception for visual object detection algorithms.2.According to the detection accuracy requirements of roadside perception,on the basis of YOLOX object detection algorithm,this thesis introduces the Rep VGG convolution block and the Coordinate attention mechanism to improve the network structure of YOLOX,and uses the Alpha-Io U loss and Vari Focal loss to optimize the loss function of YOLOX.Experiments show that the method proposed in this thesis can effectively improve the detection accuracy of YOLOX for vehicle-pedestrian objects.3.Aiming at the real-time requirements of roadside perception,this thesis designs a model compression method suitable for object detection by combining the Res Rep channel pruning and the knowledge distillation based on the attention mechanism.Experiments show that the method proposed in this thesis significantly reduces the amount of computation required by the model without a significant drop in accuracy.4.Using the intelligent roadside platform built on campus,this thesis remotely collects and annotates the actual roadside scene images to test the accuracy of the model obtained by this thesis,and deploys the model on the roadside edge computing platform with the help of model deployment tools and video streaming inference framework to test the real-time performance of the model.Experiments show that the model designed in this thesis is more suitable for roadside scenes than the vanilla YOLOX,and has certain practical application value.