Research On Algorithm Of Pedestrian Detection And Object Tracking In Intelligent Video Surveillance System

Pedestrian detection and tracking technology is an important branch of computer vision. It has a broad market prospects and value in those applications such as intelligent security, computer interaction, intelligent transportation, virtual reality and so on. Over the past decade, with the great efforts of researchers around the world, both the speed and accuracy of the technology have made great progress. However, because of the obvious different between pedestrian, unpredictable movement, changeable environment and limited computing resources, the performance of existing algorithms still cannot meet the demand of practical applications. How to increase the speed of detector without the decreasing the accuracy, as well as how to improve tracking accuracy in complex scene are still the art of research.In this paper, to solving the problems which traditional detectors may suffer from poor speed and trackers cannot update the models in complex scenes, we have made some research and innovations, which described as followed:The mainstream pedestrian detection algorithm, which using sliding windows to choose ROI for the input of feature detector, has low detection speed in monitoring. Aiming at this problem, we propose fast pedestrian detection algorithm, which is based on movement trend. Firstly, we use the information of the pedestrian movement segmentation and Kalman filter to give predictions of pedestrian position in current frames, which largely reduce the input of feature detector. Secondly, inspired by the idea of feature matching, we extract the BRISK feature descriptor, instead of pedestrian descriptor, to verify the result of first step. By this way, the algorithm both of accuracy and real-time performance have been improved. Finally, by using the characteristic of video surveillance in fixed scene as the prior knowledge, we present a novel fast method for estimating the scale of pedestrian. In the experimental part, the algorithm has been tested for four standardized video sequences. What's more, the result, which is presented in the form of DET curves and frame rate sheet, proved that our method has significantly improved the detection speed when compared with the original one, without any losing in accuracy.Aiming to solving the problem of tracking failure in consensus-based matching and tracking of keypoints algorithm (CMT), which is caused by not updating template model in the situation of occlusion, environmental changes and others, we propose a History Weighted CMT algorithm (HWCMT). In this work, firstly, a history model is constructed to weight every keypoint by appear times. Then, we adopt Gauss function as the evaluation function in history model and use it to score the importance of each keypoint contributed to the scale, angle and object center estimation. Finally, by combining the clustering and history model weighting result, our algorithm finish the pose estimation. In the experimental part, the algorithm has been tested for sixteen standardized object tracking video sequences. The experiment result indicate that our method solved the problem caused by model update in CMT algorithm and can handled part of the failure situation in original one. In brief, our improved algorithm has a definite improvement in object tracking accuracy.