Urban Road Vehicle Detection And Tracking Based On Time-space Correlation Deep Learning

With the gradual establishment of intelligent transportation system,it is highly required for vehicle detection and tracking in traffic videos.In this paper,some state-of-art methods of vehicle detection and tracking(in traffic videos)are described in details including their drawbacks.Then,a new vehicle detection and tracking method is proposed which is a combination of the MS-CNN detection network and Tracking by Detection tracking algorithm.The main work of this paper is as follows:1.In this paper,an improved version of MS-CNN detection network is proposed for vehicle detection in traffic videos.The framework is designed and implemented to tackle two problems:1)In order to solve the problem of the low accuracy in traditional moving object detection algorithm under complex environment(light mutation,picture jitters,vehicle occlusion,etc.),MS-CNN is used.By learning a great amount of vehicle training data,higher accuracy and stability in the complex conditions is achieved to detect the vehicles.2)In order to reduce the computing time of train stage and test stage of MS-CNN,PVA takes place of VGG for feature extraction in the network and better results are obtained with a more lightweight network.As a result,less time is used and the detection efficiency is improved.2.This paper improves the Tracking by Detection tracking algorithm based on the trajectory characteristics of road vehicles:1)Tracking by Detection tracking method is very suitable when using a detection network with(relatively)high detection accuracy.2)Thanks to the stability of the trajectory of the vehicle on the road,the multi-targets(vehicle)can be tracked with a linear model which is very useful under this scenery.3)Because of tracking the multi-targets at the same time,this algorithm is relatively simple and efficient.Therefore,speed is guaranteed,which meet real-time requirements.3.Based on the space-time information of the detection and tracking methods mentioned above,the detection and tracking algorithms are fused and the results were amended:1)The target of the tracking algorithm is provided by the detection algorithm the highaccuracy of which is a strong support for the tracking algorithm.2)When the target to be detected cannot be matched by tracking algorithm,the tracking target is not abandoned immediately but tracked continuously,which is very effective under the situation of temporary block.3)When the vehicle is partially blocked in which the detection algorithm cannot separate the two vehicles,the tracking algorithm can modify the results of the detection algorithm to improve the accuracy.4.This paper do experiments on the public data set called KITTI.Besides,based on the Guangdong Province highway monitoring video data which is provided by the Guangdong Provincial Communications Group,we produced a 12834 dataset with more than 60,000 positive samples of the highway traffic video vehicles for the experiments and verification of the method.In this dataset,the proposed method has a higher accuracy,faster speed and better results in comparison with the existing methods.