Research On Vehicle Detection Technology Based On Deep Learning

In recent years,due to the rapid development of our country’s modernisation process,the number of cars owned is increasing every year.The goal of the world’s leading car manufacturers to achieve smart driving technology is to reduce the occurrence of traffic accidents and ensure the safety of people’s lives and property.The low-complexity high-performance detection and tracking of targets are the primary chanllenges for intelligent driving technology.At present,radar sensor-based environment sensing devices are characterized by high complexity and high price,which are not conducive to large-scale popularization and difficult to provide semantic information of traffic scenes.However,vision-based sensors are cheap and can provide rich semantic information about the environment.Therefore,Research on the perception of the surrounding environment based on visual sensors is of great significance.Starting form the needs of road traffic application scenarios,this paper presents a comprehensive study of vehicle detection and tracking techniques and semantic segmentation of traffic scenes with a view to improving robust,detection accuracy and reducing model complexity.The main research contents are as follows.1.To address the problems that the DeepLabv3+ semantic segmentation model is prone to target discontinuity,loss of image pixel information and high model complexity,proposed an improved Deep Labv3+ semantic segmentation algorithm.In the decoder,the high-resolution low-level feature map generated by Block2 is added to alleviate the target discontinuity problem,and the direct 4-fold upsampling operation is replaced by layer-by-layer upsampling to reduce the problem of image pixel information loss;in the null space pyramid ASPP module,the original convolution is replaced by a null convolution with different combinations of null rates to expand the perceptual field of the feature map,and the depth-separable convolution is used instead of.The complexity of the model is reduced by using the standard convolution.The experimental results show that the MIOU value reaches 78.95%,an improvement of 2.75%,and the model complexity is reduced by 17.69%,which proves the effectiveness of the algorithm.2.This chapter introduces the theoretical knowledge,and studies the network structure,model training process,and target detection process of YOLOv4 target detection algorithm.Validate and analyze the performance of the YOLOv4 target detection algorithm in different traffic environments.The experimental results show that the method can successfully detect different kinds of vehicles in straight,curves and intersections under normal light,strong light and strongly changing light conditions;however,under low light and obscured conditions,missed detection is prone to occur.3.To address the problem of missed detection in the case of weak illumination and occlusion in vehicle detection,proposed an improved Deep Sort tracking algorithm based on the YOLOv4 algorithm.Firstly,the features are extracted by YOLOv4 algorithm,then estimate the vehicle’s track condition using the Kalman filtering algorithm,and finally the matching relationship between the detection frame and the prediction frame is processed using Hungarian matching algorithm.The GIOU value of the detection and prediction results is used as measurement parameters instead of IOU value.The matching performance of Deep Sort tracking algorithm is improved by it.Comparing the effect of vehicle detection with the single detection algorithm and the one after adding the tracking algorithm,the results show that there are fewer missed detections after adding the tracking algorithm,the vehicle detection effect is improved,and the robust is enhanced,and the MOTA is increased by 7.55%.