Research On Pedestrian Detection And Object Tracking Based On Video

With the development of Internet technology and computer vision, intelligent videosurveillance technology has been developed well. As the key technologies of intelligent videosurveillance, pedestrian detection and object tracking has been a hot topic in the field ofcomputer vision, which have a wide range of applications in driver assistance, intelligentmonitoring and intelligent robots. The researches of pedestrian detection and object trackingare interdisciplinary and challenging topics, which involves automation and control, computergraphics, image processing, pattern recognition and artificial intelligence etc. Many domesticand international scholars have emerged to carry out depth-research on this subject. Afterdecades of efforts, pedestrian detection and object tracking technology have made greatprogress and have proved that the method can successfully perform some specific functionsby humanlike visual analysis. However, No one algorithm can satisfy robust, high accuracy,real-time requirements simultaneously. The detection and tracking will become extremelydifficult when a target in a complex scenario. In the research background above, the depthstudy of our paper is how to detect pedestrians and track objects in complex scenarios. We tryto make some breakthroughs in pedestrian feature extraction, training and single/multi-objecttracking. The main works in this paper are as follows:First, in pedestrian feature extraction, we found that CENTRIST is a simple andreal-time feature description on the base of researches in several commonly used features.CENTRIST can achieve seamless integration with the detector. This means that CENTRISTcan be used to determine the classification of the detection window after feature extractionwithout any intermediate conversion step. However, CENTRIST only considers the relatedchanges of a pixel with its neighboring pixels, and not considers the relationship of multiplepixels among these neighboring pixels. This can lead to a bad performance in local brightnessvariation relationships. To solve this problem, we propose an improved feature descriptionbased on CENTRIST: T-CENTRIST. Compared with some other feature descriptions, ourT-CENTRIST can make up for the lack of CENTRIST and enhance the performance ofdetection. In addition, T-CENTRIST can better describe the pedestrian contour informationthan CENTRIST. Through experimental analysis, we also verified the validity of theT-CENTRIST pedestrian feature description.Second, in pedestrian feature training, we research the traditional AdaBoost algorithm,giving two problems about AdaBoost by experimentation and analysis: degradation problem and weight distribution over-fitting problem. These two problems can appear easily whengiven the training sample set contains some ‘difficult’ samples or ‘rare’ samples. Theproblems may decline the performance of the generated detector, even training to fail.Through analyzing the reasons of these two problems, this paper proposes an improvedalgorithm Adaboost, which can alleviate two problems by adjusting the distribution of thesample-weighted error. The effectiveness of the improved AdaBoost can be verified bydesigning some related experiments on some public data set. In addition, when detectingpedestrian in some complex scenarios, the occlusion problem is a common phenomenon,which may affects the performance of the generated detector. In order to solve problem ofpedestrian detection under occlusion in complex environments, we present a simple buteffective Local Area Marking Map (LAMM) algorithm, which can be used to detect thepedestrian with occlusion. After combining improved AdaBoost method and LAMMalgorithm, we construct an enhanced cascade detector, called EC-LAMM, which can improvethe detection accuracy of pedestrian detection, especially in occlusion.Third, in the single object long-term tracking, this paper analyzes two problems need tobe addressed when long-term tracking an object under a monocular camera:1) Trackingalgorithm should adapt to the changes of object with continuous deformation;2) Object mayappear or disappear in monitored scene sometimes. Tracking algorithm should determine theposition of the target in a short time and continuous tracking when the object re-appears. Inthis paper, we present a self-selective learning method to achieve single object long-termtracking. Our proposed algorithm can track a single object for a long time. Meanwhile, it canselect the appropriate samples to train an online detector according to the results of tracking.The online detector will be used to detect the tracked object by full-frame searching when itre-enter to the view. In the tracking process, we combine a simple region growing algorithmand Sobel edge extraction algorithm to get a feature point set belong to the tracked object,which is used to the Pyramidal Lucas-Kanade(PLK) tracker. In addition, in order to adapt tothe object’s shape changing, we present an on-line sample selection strategy, which willautomatically select stable tracking region to the training sample set, and update the on-linedetector by linear SVM. Such detector can be employed to find out the same object in currentframe after tracking failure.At last, in multi-target tracking, we research how to use a monocular camera formulti-object tracking and propose a novel multi-object tracking method based on WeightedIncremental Histogram Model (WIHM). Our tracking system can adapt to object’s changes ofsize and shape by WHIM algorithm. With the histogram similarity measure method:Bhattacharyya distance, the proposed method is capable of real-time multi-target trackingwith occlusion problem. Through the experiments on some public testing set and thecomparison with several state-of-art tracking methods, we demonstrate the effectiveness ofour proposed algorithm.