The Research Of Pedestrian Target Inspection And Tracking

Pedestrian target inspection and tracking is the core issue of intelligent video monitoring. The currently moving target detection algorithm is available to accurately detect the moving objects from video image subsequence. However, if there are moving background interference---not merely the interested moving targets but also the moving background objects are detected by moving target detection algorithms, to guarantee the smooth progress in the further target features extraction and target tracking, pedestrian target must been detected precisely on the phase of pedestrian detection when detect and track the pedestrian. Camshift algorithm is the tracking algorithm based on targets’ color information, but if the target is obscured, the targets’ color will miss. Consequently, the algorithm will be unable to track the target. The paper mainly researches the above issue of pedestrian detection and tracking, 3 aspects of the contents are listed as below:Offline-training pedestrian classifier: Because the university of the pedestrian detection classifier included by Opencv is limited, the classifier cannot work in diverse tracking scene. This paper studies to dig the positive pedestrian samples and negative background samples out of the obtained video image subsequence in special tracking scene, extract the Haar feature of samples’ image, then train the positive and negative samples with the Adaboost algorithm, getting the pedestrian detection classifier that is applicable for special tracking scene.In respect of pedestrian detection, objects, detected by three frames difference algorithm from the tracking scene, include cars, shaking leaves and other non-pedestrian objectives as well as pedestrians. To detect the pedestrian targets precisely, firstly get the difference image by three frames difference algorithm, then process the growing of seeds’ zone for this difference image, filling up the cavity in the moving and changing zone of the frame difference image, obtain the relatively completed moving targets and the outer rectangle of the moving target’s outline, upload the pedestrian detection classifier obtained by training in tangle moving zone, realize the precise detection of pedestrian target at the end.In respect of pedestrian tracking, in view that Camshift algorithm couldn’t track when target been occluded and even miss the tracking target, we introduce the Kalman filtering algorithm into the tracking frame. Initially, divide the pedestrian target out of moving objects by pedestrian detection algorithm improved in this paper. When pedestrian are occluded, Kalman filter will predict the position of the current frame according to the move status of the target in the last frame. The Camshift algorithm takes the predicted position as the initial position, then processes mean shift iterations until meet the certain convergence requirements. Finally, alter the Kalman filter parameters and move status of the target with the result of Camshift algorithm as measured value. Thus the continuous target tracking can be achieved through circulation of the above process. This paper introduces the improved algorithm into the multiple pedestrian tracking, by experiments the paper proves that the multiple targets tracking algorithm also can achieve preferable tracking effect when moving targets occluded by each other and meet the requirement of timeliness at the same time.