Research On Pedestrian Tracking Algorithm Based On RGBD Video Sequence

Pedestrian tracking algorithm based on video sequences is one of the hot research topic in the field of computer vision,the algorithm research is applied extensively in the field of intelligent monitoring,also has important role on pattern recognition and behavior recognition of other areas of computer vision.Pedestrian tracking algorithm consists of a pedestrian detection and pedestrian tracking research,the pedestrian tracking algorithm based on video sequence is to detect the pedestrians target in video sequences,and then to track the detected pedestrians.At present,there has been many research scholars made a lot of research achievements in the field of pedestrian tracking algorithm,and proposed all kinds of good pedestrian tracking theory and algorithm,each one has its own advantage and disadvantage.As the rapid development and wide application of the Kinect camera,there has been more and more researchers begin to use the superior performance of Kinect camera to further improve the tracking accuracy of pedestrian tracking algorithm.Considering the superior performance of the deep learning in the fields of computer vision,this paper proposes a framework based on the deep learning and RGBD video sequence of pedestrian tracking algorithm in this paper.The main research contents of this paper include:(1)Put forward the deep learning framework of Faster-RCNN as the pedestrian detection model.The framework has the best target detection effect in deep learning,through RCNN and Fast RCNN development,Faster-RCNN is united the four basic steps of the target detection into a deep web framework,composed by regional network and Fast RCNN network,to extract the characteristics of pedestrians candidate,and refine the pedestrians location,on the basis of the optimization calculation complexity,not only can improve the speed,but also could improve the accuracy of the pedestrian detection.(2)Use the person re-identification algorithm for the pedestrian tracking algorithm.The traditional pedestrian tracking algorithm is commonly used the particle filter and kalman filter method,and so on.Instead of using the traditional tracking methods,this paper chooses the person re-identification algorithm based on the deep learning framework.The algorithm use CNN to extract the detected pedestrians feature,compare and match the detected pedestrians between several frames,to complete the pedestrian tracking step by computing cosine similarity.Through the experiment results show that the proposed tracking algorithm has obviously improved the tracking accuracy than before tracking algorithm.(3)Using RGB video sequence to complete the pedestrian detection and pedestrian tracking.This article chooses the RGB video sequences separate out from the RGBD dataset for pedestrian detection and tracking,compared with the standard dataset the pedestrian tracking accuracy,and compared with the before algorithm accuracy.The experiment results show that using RGB video sequence single based on deep learning framework for pedestrian tracking,the tracking accuracy algorithm has been obviously improved than before.(4)On the basis of using RGB video sequences,and further adopt RGBD video sequences for pedestrian detection and pedestrian tracking.Input the RGB and depth image into the system at the same time,to extract characteristics of the RGB and depth images at the same time,through the feature of the depth image to make up for the feature of the RGB image.Experimental results shows that the tracking accuracy of using the RGBD video sequences for pedestrian tracking is higher than using single RGB video sequence to track.