Research Of Vehicle Detection Technology Based On Deep Convolutional Network

Vehicle detection is an important research directions in the field of object detection.For the problems of poor adaptation in different environment and easy to be disturbed by some extraneous factors,this thesis studies the vehicle detection issue from two aspects: the artificial extraction feature and the automatic extraction feature based on deep convolutional network.In addition,in order to improve practicability and portability,a vehicle detection system is established with Android development software.The specific research work is as follows:1.Building the image dataset.The image dataset contains 1500 vehicle images,which are obtained by two ways,one is searching on the Internet,and another is getting from GTSDB(the Traffic Sign Detection Benchmark).All the images are taken under different illumination and different angles as far as possible.2.Vehicle detection based on the artificial feature extraction.As an artificial landscape,vehicles have obvious shape characteristics.This thesis employs HOG and Haar-like as shape extraction feature algorithms,and SVM and Adaboost as learning algorithms.The candidate regions are generated with the sliding window to complete the vehicle detection.After testing,the average accuracy of the two methods is 60.76% and 74.31% respectively.3.Vehicle detection based on deep convolutional network.This thesis selects the Faster-rcnn detection framework,a representative detection framework based on region proposal.In the experiments,VGG+RPN and ZF+RPN are employed to train the model;the average accuracy of the two methods is reach 86.33% and 85.83% respectively.4.Vehicle detection system based on Android.This system is developed with Android Studio and eclipse development platform and realizes asynchronous communication between mobile application and server based on Http communicating protocol.The mobile application gets images by capturing or choosing from albums,and then uploads the images to the PC server.The server receives the images from the client,detects the vehicles in the images and sends the results back to the mobile applications.After testing with 300 images,the average transmission speed of system is about 1 second.