Pedestrian Tracking And Pose Estimation In Surveillance Video

With technology improvement in computer science and image processing and increased public safety consciousness, video surveillance system has been widely used in the past decade. Video synopsis is an effective method to obtain useful information from massive raw videos, and its basic foundation is object detection and tracking. Pedestrian detection and pose estimation mean locating the full body or body part in image or video, are some of the hottest and most difficult research areas in computer vision.In this thesis, a survey on traditional pedestrian detection and pose estimation was conducted. Existing tracking algorithms are not ideal for pedestrian occlusion situations, especially in the surveillance video with uncalibrated monocular camera setting losing depth information. Meanwhile, traditional pose estimation methods rarely utilize the appearance similarity in adjacent poses and spatio-temporal continuity in pedestrian tracks. Based on these, the main works are as follows.One is proposed a multi-objects pedestrian tracking method based on motion and appearance information fusion. First, motion and appearance distance measures are calculated between each detected pedestrian object and each track. Then the weight coefficients are adjusted according to the distribution of the calculated distances, so that all the distances are fused to a single cost value, which forms the cost matrix. Finally, Hungarian algorithm in assign problem is employed to associate each pedestrian object to its track using the cost matrix above. Thus, the effect of multi-objects pedestrian tracking in monocular video when temporally occlusion occurred is improved. Tracking results in multi dataset videos indicate the feasibility and effectiveness of the proposed algorithm in temporary occlusion situations.The other is proposed a human orientation pose estimation method based on soft label and spatiotemporal constraints. First, label distribution learning method is applied to make use of the samples of adjacent pose labels to generate a soft label pose estimator. Then implement the estimator on the pedestrian track sequence acquired above to form an estimated pose sequence on time domain. Finally, hidden Markov model is used to model the estimated pose sequence, the velocity in pedestrian track adaptively constrains the pose to obtain a reliable estimation. The continuity in pedestrian orientation pose itself and spatio-temporal continuity in the pedestrian track are exploited to enhance the sequential consistency in pose sequence. Compared to mainstream classifier, the proposed method can generate orientation pose of a better sequential consistency in the pedestrian sequence.