A novel method for tracking and counting pedestrians in real-timeusing a single camera

摘要

This paper presents a real-time system for pedestrian tracking innsequences of grayscale images acquired by a stationary camera. Thenobjective is to integrate this system with a traffic control applicationnsuch as a pedestrian control scheme at intersections. The proposednapproach can also be used to detect and track humans in front ofnvehicles. Furthermore, the proposed schemes can be employed for thendetection of several diverse traffic objects of interest (vehicles,nbicycles, etc.) The system outputs the spatio-temporal coordinates ofneach pedestrian during the period the pedestrian is in the scene.nProcessing is done at three levels: raw images, blobs, and pedestrians.nBlob tracking is modeled as a graph optimization problem. Pedestriansnare modeled as rectangular patches with a certain dynamic behavior.nKalman filtering is used to estimate pedestrian parameters. The systemnwas implemented on a Datacube MaxVideo 20 equipped with a DatacubenMax860 and was able to achieve a peak performance of over 30 frames pernsecond. Experimental results based on indoor and outdoor scenesndemonstrated the system s robustness under many difficult situationsnsuch as partial or full occlusions of pedestrians