Research On Pedestrian Tracking Algorithm Based On Particle Filter

Object tracking has a wide range of applications in video surveillance, intelligent human-computer interaction, visual navigation of robots, intelligent transportation, behavioral analysis and medical diagnosis, etc. Pedestrian is the main goal of tracking in most scenes of object tracking. Pedestrian tracking has important research significance and application value in object tracking. However, pedestrian tracking in complex environment is still facing many problems due to the arbitrariness of human motion, illumination and change of pedestrian postures, complex background and occlusion. Existing pedestrian tracking algorithms lack observation model of good robustness in describing the features to adapt to a variety of complex environments, thus they lack applicability in the real world and cannot achieve accurate robust tracking; on the other hand, the research in multiple pedestrians tracking algorithm is comparatively few. Most of the existing algorithms can achieve pedestrian tracking in the static background, few of them can achieve pedestrian tracking in the moving background; In addition, the current pedestrian tracking research often uses data of single video, which has shortcomings in small tracking field of view, less information, mutual occlusion due to the visual angle, etc, and thus has difficulties in achieving accurate tracking of all pedestrians in pedestrian intensive scenes.Since particle filter is able to handle any non-linear and non-Gaussian distribution system which is more accurate in describing tracking problems in actual scenes, the dissertation puts forward pedestrian tracking algorithm based on particle filter theory in respond to the above issues. The main research content includes the following aspects:1) Pedestrian tracking in video sequences based on multi-feature particle filter of simulated annealing and particle swarm is proposed based on the particle sampling optimized particle filter theory and the improved robustness of the observation model by multiple features fusion. Starting from the single pedestrian tracking problem, this dissertation uses simulated annealing and particle swarm algorithm to optimize particle sampling results based on the similarity between particle swarm and particle filter as well as the improved global extreme conditions of particle swarm by simulated annealing to deal with the particle diversity scarcity phenomenon in particle filter algorithm; then to solve the problem of single pedestrian tracking in video sequences, the state-space model is improved to enhance the particle’s tracking capability of objects; in view of the limited ability of object recognition of the single pedestrian feature observation model under the influence of complex environment, background and noise interference factors, the dissertation combines with three types of feature information which have complementarities between each other and adjusts feature weights self-adaptively in designing the observation model to increase the object identification of algorithm in observations. Compared with pedestrian tracking using single feature, the proposed algorithm improves the accuracy and stability in tracking and can achieve better tracking results in the case of object translation, posture change, complex background, partial occlusion, etc.2) Multi-pedestrian tracking algorithm in video sequences based on discriminative model is proposed by introducing bag of features algorithm into pedestrian tracking in both static and moving background which transforms a large number of feature sets into a smaller number of feature dictionaries and bags to create a discriminative model, and effectively solves complexity problem of multi-pedestrian feature extraction. Multi-pedestrian tracking is more complex than single pedestrian tracking, and using multiple features in multi-pedestrian tracking will cause a large amount of calculations, therefore it is not suitable to solve the problems of multi-pedestrian tracking. In view of this, the dissertation introduces a simple-calculating, low-complexity bag of features algorithm, and combines with superpixel and the extraction of LBP block features jointly to establish a discriminative model to determine the pedestrians in video sequences. Unlike conventional tracking algorithms, multi-pedestrian tracking algorithm in this dissertation proposes two detection methods which can be applied to both the static and moving background in the detection phase, thus improves the applicability of the algorithm; meanwhile, the problem of pedestrians’ mutual occlusion in multi-pedestrian tracking process is processed in order to prevent the drift and lost phenomenon caused by the object block. Experimental results show that the proposed tracking algorithm has better stability and robustness in processing the objects’ translation, occlusion, the interference among pedestrians, illumination, change in walking speed as well as analogue interference, etc.3) For the problem of occlusion among multiple pedestrians in pedestrian intensive scenes, multi-pedestrian tracking algorithm based on video sequences and laser point clouds is proposed by combining feature information contained in video sequences and three-dimensional information of laser point clouds. It is difficult to accurately track all of the pedestrians by using data from single video in comparatively intensive scenes for the lack of tracking information caused by pedestrian occlusion. The algorithm first obtains classification parameters with better detection performance through interpreting, classifying and detecting laser point data to separate the data of pedestrians and vehicles, then achieves the fusion of video sequences and laser point clouds by the method of calibration, analyzes the region of interest after fusion by using the complementary between the two data; associates particle state with the detection by using a combination of pedestrian detection results with particle filter tracking algorithm; and processes the disappearance and appearance of pedestrian objects. As a result, pedestrian tracking in comparatively intensive scenes has been achieved. Experimental results show that the fusion of two data improves the performance of pedestrian detection and multi-pedestrian tracking algorithm based on that is able to handle pedestrians’ mutual occlusion, object appearance, etc, thus has better tracking results.