A Study Of Multi-pedestrian Motion Detecting And Tracking Under Complex Background

Multiple pedestrian target detection and tracking technology has important research value and application value in many areas of intelligent video surveillance, intelligent transportation fields and modern defense. Due to the complexity of the scene and pedestrian target variability in a complex background scene, study issued by the robustness and reliability of video tracking algorithm is facing enormous challenges. Based on the previous research, this paper try to summarize more pedestrian tracking algorithm and find relatively outstanding algorithm and make appropriate improvements to develop a multi-pedestrian tracking system.In this paper, starting with the characteristics of the target pedestrians, some common feature extraction method has been researched. Feature extraction contributes to locate the tracked object within the image and is the basis for pedestrian detection based on a static image. By comparison test, several pedestrians feature extraction methods are summarized and evaluated, proposing the use of combined LBP-Haar features and HOG-LBP features of extraction algorithm and corresponding pedestrian detection strategies. Each image sequence using two strata of feature extraction algorithm, experiments show that the method for pedestrian detection based on static images shows excellent. For the detection based on dynamic frame, this article focuses on three background subtraction method in which the Gaussian mixture background modeling is the classical one in view of its excellent detection results. The second is the codebook modeling algorithm, different from the previous codebook improved herein retain codebook model in the RGB color space and RGB-YUV color space only when luminance conversion. The advantage is improving the speed from the squaring to linear transformation and retaining foreground object color appearance as to track the stage with the appearance of the characteristic color for target identification part. The third is an adaptive pixel-based segmentation algorithm, which is a relatively new non-parametric background modeling method, the principle straightforward and easy to implement, have significantly improved their performance. Comparative experiments on three methods show the performance of detecting moving foreground.For target tracking section, the paper focuses tracking algorithm based on data associ-ation. This paper studies two ways in which the selection, the first method is based on a real-time algorithm for people tracking using contextual reasoning, which is used Finite State Automation and determines the target current state based previously saved state and its current characteristics. The second method is based on discrete-continuous optimization to make the analysis of trajectory assumptions by minimizing the energy function. Both can solve target split, merge,entering or leaving the scene and other issues.Finally, this paper presents a set of multi-pedestrian motion detection and tracking system. The tracking system uses the idea of two strata of feature extraction algorithm, based on codebook modeling and discrete-continuous optimization with FSA forming a tracking system under the complex background, through experiments on the accuracy and robustness for verification.