Anti-occlusion Vehicle Detection Based On Multi-scale Parallel Network

With the rapid development of information technology nowadays,the traditional transportation technology and means have not adapted the requirements of social development,intelligent transportation system has been paid more and more attention.As the “intelligence agent” in the intelligent transportation system,vehicle detection system has been widely used in traffic monitoring,vehicle management,electronic toll collection system and other scenarios due to its features of real-time monitoring of traffic conditions and obtaining vehicle information.This paper focused on the efficient vehicle detection algorithm and software application system in the intelligent transportation system.The main research work and corresponding results of this paper are shown as follows.(1)Aiming at the problem of insufficient image data for training vehicle detection network model,multiple data expansion methods for vehicle image were studied.Firstly,using the advantage of generating confrontation network in unsupervised learning,the self-collected vehicle images were processed to generate a large number of transformed vehicle images.Secondly,the on-line occluded vehicle image expansion technology proposed in this paper was used to expand the vehicle image online.Through adaptive occlusion of the input image,the occluded vehicle image was generated.Finally,the self-built data set of this paper was constructed by combining the extended vehicle images obtained by different methods with the KITTI data set and some vehicle images in the domestic open data set.(2)Aiming at the problems of low detection accuracy of small size vehicle,missed detection of occluded vehicles and false detection of background area in traditional vehicle detection algorithms,a multi-scale parallel anti-occlusion vehicle detection method based on attention mechanism was proposed.Firstly,we used the method of multi-scale feature extraction to get the regional candidate boxes with different scales.Secondly,feature information of different scales was fused.Then,the fused feature information was input into the target detection module for target classification and accurate border regression.Finally,the detection performance of the whole network for occluded vehicles was improved by adding an occlusion module consisting of an on-line occluded vehicle image expansion network.Through a large number of data tests,it was found that the vehicle detection method proposed in this paper was superior to other comparison methods.(3)Aiming at the requirement of intelligent transportation system for the engineering of vehicle detection algorithm,the design and implementation of vehicle automatic detection and recognition software system is carried out.By encapsulating and modules separating of the proposed algorithm in this paper,and embedding the reserved API interface into the programming software interface,the final software system could achieve high-precision real-time detection of the target vehicle image and the output of the parameters information of the vehicle target.