Road Vehicle Detection Algorithms Based On Deep Learning

In the natural environment,the traffic scene where the vehicle is located is complex,and there are problems such as different scales,overlapping and ambiguous vehicle targets.This paper combines vehicle detection and deep learning,through a variety of models for targeted optimization and improvement,better achieve the road vehicle target accurate recognition.The main research content is as follows.To address the problems of dense detection targets and different scales of vehicle targets in the process of vehicle detection,which leads to insufficient extraction of vehicle information in the detection process.This paper proposes a road vehicle target detection algorithm based on the cascade structure of yolov5-based multi-module attention mechanism.Firstly,a new attention module is proposed by fusing the advantages of the convolutional attention module CBAM and the efficient channel attention network ECA-Net,and adding them to the network structure of the Yolov5 algorithm to focus the detection points in the target detection process by judging their importance on the key regions.The multi-scale detection feature map is then redesigned with a larger scale feature map to capture the more shallow feature information that is easily lost.Following by replacing the complex Bounding box loss function in the output layer with Efficiency Intersection over Union instead,thus speeding up the convergence and improving the localization accuracy is improved.Experimental results demonstrate that the improved Yolov5 algorithm improves the mean detection accuracy(m AP)from 91.3% to 96.3% compared to the original Yolov5 algorithm,and has higher detection accuracy and faster inference speed on the KITTI dataset compared to the classical single-stage detection algorithm SSD and the two-stage detection algorithm Faster R-CNN.To address the problems of large number of model parameters and slow detection speed of vehicle detection algorithms,this paper analyses the lightweight network-related algorithms such as deep separable convolution,GSConv module and Ghostnet module,and applies the GSConv module and Ghostnet module to the backbone network part of the multi-scale attention mechanism Yolov5 algorithm and Neck.Finally,the balanced performance of accuracy and real-time between different lightweight modules is tested on the KITTI dataset.The experimental results demonstrate that the improved algorithm in this paper is able to perform the task of vehicle detection faster while maintaining a high accuracy rate.The improved vehicle detection model designed in this paper based on the Yolov5 algorithm has high vehicle detection accuracy and fast detection speed,and can meet the vehicle detection tasks in a variety of scenarios,which can provide data information for further research work on license plate recognition,vehicle tracking and traffic statistics,and has strong practical application value.