Vehicle Detection Method In Natural Scene Based On Convolutional Neural Network

The detection for vehicles in natural scenes is an essential part of autopilot technology and intelligent transportation system,and a prerequisite for license plate recognition and vehicle type recognition,which has very important research significance and practical value.Traditional vehicle detection methods are mainly based on feature extraction and classification of image.They rely too much on the shallow features selected manually,which can greatly reduce efficiency and accuracy in complex and unfamiliar environments.Therefore,traditional methods exist great limitations and are difficult to optimize.In recent years,deep learning theory has been improved continuously,and its ability to automatically extract features has greatly reduced the difficulty of solving problems mentioned above,so as to achieve higher detection accuracy.This paper takes the deep learning knowledge as the theoretical support,and explores the two-stage vehicle detection framework based on fast proposals generation and object recognition network.Due to the limitations of existing vehicle detection algorithms,we design a method to detect vehicles under natural traffic scenes.The proposed method mainly carried out the following work:Firstly,aiming at the problem that many proposals generation algorithms are too slow and have low accuracy of positioning,this paper proposes a reBING-MLBP algorithm to quickly generate a series of proposals with ranking scores which probably contain vehicles.The reBING-MLBP combines advantages of gradient and LBP features,and optimizes proposals using object's edge and superpixel fusion techniques.It can generate 1000 high-quality proposals within 12ms,reaching a DR of 96.2%and a MABO of 79.3%.Therefore,reBING-MLBP can solve the problem that speed of vehicle detection algorithms is limited by performance bottleneck of proposals generation algorithms.Furthermore,we combined reBING-MLBP with the CNN network to realize a vehicle detection method.Experiments were performed on the VOC2007 dataset and the KITTI dataset.This method has achieved a higher detection rate and are faster than Fast R-CNN detection algorithm.Secondly,in order to achieve more accurate positioning and identification of vehicles in complex and variable traffic scenes,we realize a vehicle detection method based on the feature pyramid structure using multi-scale CNN network.This method extracts CNN features with a ResNet-50 network,which has simpler structure and deeper layers.For proposals generation,a RPN network is applied into the method,so it finally becomes an end-to-end detection network.Besides,the feature pyramid structure is introduced for multi-scale information,and the RoI Align strategy for region mapping pooling is introduced too,leading to improved detection rate of small-scale vehicles.Experiments and relative analysis were performed on the KITTI Car dataset.This method can reach a AP of 88.88%when the overlap threshold is limited in 0.7,with the computing speed of 0.4s per image.Therefore,it is valuable and can be applied in complex natural scenes and traffic environments.