Research On Improved Yolov5 Vehicle Detection Algorithm Based On Convolutional Neural Network

As the deep learning advances,the technology of target detection has been widely applied into various fields,such as intelligent transportation,driverless field of automobile,etc.Among them,vehicle detection has achieved a relatively mature application in the fields of intelligent transportation and military.However,the current vehicle detection algorithm still has challenges.The difficulties mainly include the imbalance of data samples in the data set used to train the target detection network,which makes the trained weight file ineffective in the actual vehicle detection;Due to the complexity of training data set and actual road background,different time,scene and weather lead to poor feature extraction effect of target detection network,which affects the detection effect;Due to the large number of small target vehicles in the data set,it is difficult for the target detection network to detect them all,which will affect the final detection performance.In order to solve the above problems and improve the performance of vehicle detection algorithm,this paper is based on the single-stage target detection algorithm yolov5 of convolutional neural network.The main research work is as follows:(1)This paper will use the single-stage target detection network yolov5 as the framework of vehicle detection algorithm,aiming at detecting bus,car and truck.Caused by the imbalance between the date set and actual situations,some mistakes of missed detection and false detection will occur when the target detection algorithm is trained and tested.In such a case,the performance of the target detection algorithm will be affected a lot.This paper uses the deep convolution generation countermeasure network dcgan to generate vehicles with less data samples so as to reduce the unbalanced proportion of data samples,so as to improve the data set used at the input of yolov5.Experiments show that using dcgan as image generation improves the input of yolov5 and improves the effect of the algorithm on vehicle detection.(2)Due to different weather,time and scene changes in the actual situation and data set,it will be difficult to accurately extract vehicle feature information under this complex background,which will affect the performance of target detection algorithm.This paper adds the attention mechanism ECA module to the original yolov5 feature extraction network.This module is a lightweight channel attention module,which can change the weight values of different channels during feature extraction through cross-channel interaction,effectively improve the feature weight ratio of vehicle targets,and then improve the feature extraction ability of the model.Experiments show that adding the attention mechanism ECA module to the feature extraction network of yolov5 can effectively improve the overall performance of the vehicle detection system.(3)In the training data set,because there are many small target vehicles in the picture,it is difficult for the target detection network to detect them all,which affects the overall performance of the detection system.Aiming to develop the ability of detecting small target vehicles,this paper attempts to perfect YOLOv5’s feature fusion network and strengthens the feature fusion network of target detection network towards small-sized vehicles.Experiments indicates that the modified YOLOv5 algorithm can better ameliorate the detection performance of the system towards small target vehicles when training and testing on the BDD100 K driving video data set.