Research On Vehicle Detection Algorithm Based On Deep Convolutional Neural Network

In recent years,autonomous driving research and smart transportation construction have advanced by leaps and bounds,and vehicle detection technology has become a research hotspot in the industry.The deep convolutional neural network has certain characteristics of rotation and translation invariance,and is widely used in vehicle detection tasks.Among them,the YOLOv3(You Only Look Once version 3)algorithm is one of the main detection algorithms at present.However,in vehicle detection applications,the YOLOv3 algorithm has the following problems:(1)YOLOv3's a priori detection frame generation mechanism does not take into account the significant differences in the size of different types of vehicles,resulting in the a priori detection frame can not fit the real target well The size reduces the detection accuracy;(2)With the deepening of the network,the feature extraction network of the YOLOv3 algorithm will produce a certain degree of information loss,affecting the detection accuracy of the deep network;(3)The deep detection feature map of the YOLOv3 algorithm is too small,Is not conducive to the accurate detection of large and mediumsized targets.In response to the above problems,this paper proposes an improved YOLOv3-based vehicle detection algorithm IAVD-YOLOv3(Improved Algorithm for Vehicle Detection based on YOLOv3),on the premise of ensuring real-time,effectively improve the vehicle detection accuracy.The main work of the paper is as follows:(1)First of all,a new anchor-Box size generation strategy is proposed.This strategy uses the K-means++ algorithm to cluster the real target box size,and uses the squared Euclidean distance as the distance metric between samples.It overcomes the problem that the original method clustering results are unstable and easy to fall into the local minimum,while enhancing The distinction between different types of vehicles is improved,and the detection accuracy is improved.Then,in the case of fewer small targets and more medium and large targets in the vehicle detection task,by increasing the equivalent sampling step of the deepest feature map,the effective receptive field of the network is expanded,so that the network can extract features from a larger range,In order to adapt to the large size target in the data set.Then,aiming at the problem of loss of information in the original YOLOv3 feature extraction network,drawing on the idea of Dense Net,a new feature extraction network is designed,and dense modules and transition modules are added to the network,so that shallow features can be more fully transferred to the deep layer of the network.Finally,five detection layers were redesigned.The five detection layers together form a feature pyramid to perform vehicle detection tasks,enhancing the detection capabilities for vehicle targets,especially medium and large-sized vehicle targets.(2)Based on GTX1080 Ti graphics card,KITTI automatic driving dataset and self-made vehicle detection dataset are used to evaluate model performance.We use average accuracy and detection frames per second as evaluation indicators.The algorithm of this paper is compared with YOLOv3,SSD,Faster-RCNN.First,based on the KITTI dataset,verify the influence of the prior detection frame clustering method on the accuracy of the vehicle detection algorithm.Then,the algorithm of this paper is compared with the above three algorithms to test and evaluate the pictures in the KITTI dataset from the two aspects of average accuracy and running speed.Finally,the self-made vehicle detection dataset is used to evaluate the algorithm in this paper.Through experiments,it is proved that the average accuracy(mAP)of the IAVD-YOLOv3 algorithm is higher than that of the YOLOv3 and SSD algorithms,which is close to the Faster-RCNN algorithm that does not meet the real-time requirements.Futher,the IAVDYOLOv3 algorithm can better adapt to vehicle detection tasks in new scenarios.