Research On Vehicle Detection And Recognition Method In Camera Network

In recent years,drived by the continuous development of video surveillance network and high-definition technology,the quantity of the traffic surveillance video id growing at a rate of 55%every year.Using computer vision technology to deal with traffic video to get effective information should be gradually being taken seriously.For the camera network,not only needs to process video under the single camera,but also needs to get the relationship of multi-camera,so as to get the continuous tracking for one vehicle in the cameras' network.In the paper,the Faster Region with Convolutional Neural Network feature is used to detect vehicles.We using the method of fine tuning to train the vehicle detection model.Then,a vehicle tracking algorithm based on overlapping ratio is proposed under the single camera.Whether the vehicle belongs to the same trajectory according to the overlapping ratio of the target vehicle in the adjacent two frames.Firstly,the RCNN,Fast R-CNN and Faster R-CNN are introduced in detail.In order to improve the accuracy of detection,this paper using Faster R-CNN network training the model of vehicle detection and build a good foundation for vehicle tracking.The vehicle occlusion compensation algorithm based on kalman filter is proposed for vehicle occlusion,failure of detection and the missed detection.Based on the prediction mechanism of kalman filter,after calculating the overlapping ratio,if there are trackers or vehicles failing to match successfully,kalman filter is used to predict the location of vehicle for matching,so as to avoid the failure of the vehicle matching.In addition,the results are visualized using the homography.For the videos of multi-camera,the light,shoot angle and background are different.All of this increase the vehicle re-identity difficult.In this paper,the similarity measure formula of vehicle re-recognition is established according to the time-space relationship,the vehicle type attribute and the vehicle Convolutional Neural Network characteristic.The spatial information of the camera can be obtained by using the geographical position of the camera.The probability distribution of the transit time probability between the cameras can be obtained from the video statistics.In addition,the vehicle type classification model is obtained by retraining using the GoogleNet.And CNN feature could get from VggNet.Then it could calculate the similarity of vehicles from different cameras to finish the vehicle re-recognition.