Research on Video Sequence-based Detection of Pedestrian Flow
|Keywords||head detection hair-color model non-parameter probability hough transformation meanshift|
At present, pedestrian flow measurement raises focus of researchers in the field of machine intelligence. Pedestrian flow measurement is widely utilized in the application of transportation monitor and control, intelligence vehicle, automatic passenger counting and surveillance of banks, supermarket and museum. The information retrieved from measurement consist the essential source of elements in terms of efficient management. Nowadays, the pedestrian flow measurement instruments with infrared ray or pressure sensor failed to meet the requirement of high accuracy because of the disadvantage of low measurement accuracy. However the computer vision technology is capable of establishing an intelligence system with assistance of image processing. In consequence, the computer vision technology is of foundation in raising precision of pedestrian flow measurement through video sequence capturing by camera, motion object location, recognition and tracking.With the precondition of introduction of pedestrian flow measurement technology and comparison of existing study status, the algorithm framework aiming at pedestrian flow measurement is proposed with digital image processing technology. The algorithm presented consists of multi-object detection and tracking, both of which are detailed in two chapters.The paper presents a serial of detection algorithms step by step and establishes a novel multi-object recognition approach. Firstly, pre-detection of head with fast gradient hough transformation is implemented; secondly, hair-color classification is proposed to filter the candidate targets through modeling hair-color distribution; finally, introducing non-parameter probability theory, the probability of circle existence model is established, which finalizes the stages of head detection by locating the head. Experimental results indicate that the head detection algorithm is more accurate and robust compare to detection results of gradient hough transformation detection.In terms of motion object tracking, the paper proposes a meanshift-based object tracking algorithm to execute the tracking task. Meanshift features with robust and low sensitivity to changes of tracking target contour for the reason that it refers to color information and non-parameter probability model. Furthermore, Kalman filter is induced into meanshift tracking algorithm to solve the failure of object tracking when object moving too fast. Finally, a pedestrian flow measurement technology combining the multi-object detection and tracking algorithm is constructed in the paper. Moreover, video sequences of both scenes of bus and entrance are utilized as experimental data through implementation of object recognition, tracking and counting. The experimental results certify the availability and robust of the measurement algorithm compared to existing approach.