Research On Multi-target Vehicle Detection Based On Deep Learning

Vehicle detection is a practical application field for traditional target detection and has broad application prospects in the fields of surveillance,driverless and intelligent transportation.The traditional vehicle detection algorithm is limited by the types of extracted features,is susceptible to external factors of the environment,and has a large dependence on artificial,which makes the generalization ability of the algorithm poor and the detection efficiency low.Based on the deep learning vehicle detection algorithm,due to the self-learning ability,the network can learn more feature information through a large number of data sets and has better detection performance.The use of deep learning networks for vehicle detection research is of great significance.Since the YOLOv2 algorithm has better real-time performance than other deep vehicle detection models,this paper uses YOLOv2 as the research entry point.The main research contents of this paper are as follows:(1)For complex and harsh environments,we use the Retinex algorithm to enhance the image and then use median filtering to denoising the enhanced image to improve the detection of the model in harsh environments.In order to improve the detection efficiency of the model,the paper have added a discriminating module for the blurred image.If the image is blurred,the input image is enhanced,otherwise,it is directly input into the model for detection.(2)Although YOLOv2 has good real-time performance,YOLOv2 still has two problems in the scene of vehicle detection:(1)Some repeated convolutional layers of the YOLOv2 model are not very effective for vehicle detection with high similarity of class,and will increase the amount of calculation;(2)The increase of the number of convolution layers and the pooling operation will cause some small target object features extracted by the model to be lost in the network model,which reduces the detection accuracy.In this paper,the improved model YOLO-DV of YOLOv2 is proposed.This model reduces the calculation parameters of YOLOv2.By splicing the characteristics of the shallow layer of the network with the features of the deep layer,the fusion between the multi-layer features is realized,and it is difficult to improve the small target.Detect the detection efficiency of the sample.In the vehicle dataset of Pascal VOC2007/2012,the YOLO-DV model increased mAP by 1.3 percentage points compared to YOLOv2.(3)During the training of YOLOv2 network,most of the trained samples are negative samples that are easy to classify,and a small part is difficult to classify samples,resulting in negative samples that are easy to classify and contribute too much to the loss of YOLOv2.The FocalLoss loss function can weaken the contribution of easily classified samples to loss,thereby increasing the impact of difficult samples on loss.This paper uses the FocalLoss loss function to improve the class loss function of YOLOv2.Using this improved loss function can improve the detection accuracy of the model in dense and difficult to detect scenes.Both YOLOv2 and YOLO-DV with improved loss function have improved recall and mAP.(4)In order to further verify the effectiveness of YOLO-DV in vehicle detection,this paper compares and tests the detection results of YOLOv2 and YOLO-DV in the data sets of three road conditions: unobstructed,generally unobstructed and blocked.Under the data set of the general unobstructed scene,the YOLO-DV model increased the precision rate by 2% and the mAP by 1.7 percentage points.