Study Of Visual Based Pedestrian Detection And Tracking Algorithm

In the various applications of computer vision, the object detection and tracking is one ofthe most important and basic tasks. Some of perspective applications include automatic driv-ing system, robot control, video compressing, visual based control, motion recognition, humanmachine interface, medical imaging, augmented reality, and visual based intelligent surveillancesystem. Although the object detection and tracking has been studied for more than ten years incomputer vision community, it is still an active research area. In present, there is not an objectdetection and tracking system which is general, robust, accurate, efficient, and real time. Becausethe human body is non-rigid, the scenes are cluttered, and there are a lot of interactions amongthe pedestrians or between the pedestrians and the scenes, the human detection and tracking isone of the most difficult challenges among those tasks mentioned above.The highlights and main contributions of the dissertation include:1. A Novel Pedestrian Detection Method in Natural ScenesA pedestrian detection algorithm is presented, which use nested cascade of Look-Up Table(LUT) Gentle AdaBoost with the Edge Orientation Histogram (EOH) features and extendedHaar-like features. In this pedestrian detection method, the node learning algorithm based on thecascade risk cost function is adopted in training cascade architecture to associate the performanceof each stage classifier with the overall performance of cascade classifier and improve detectionperformance. Experimental results demonstrate the robustness and efficiency of the proposedframework. It proved itself to be both accurate and fast. It can also distinguish pedestrian fromthe other moving objects, such as moving car.2. Feature Transformation and SVM Based Hierarchical Pedestrian Detection withMonocular Moving CameraIn environments where a camera is mounted on a freely moving platform, e.g. a vehicle,pedestrian detection becomes much more difficult. Especially, in cluttered scenes, pedestriandetection is more challenging. In this paper, we propose a coarse-to-fine pedestrian detectionbased on feature transformation and SVM with a monocular moving camera. Firstly, a coarsepedestrian detector is learnt by Look-Up Table(LUT) Gentle AdaBoost cascade. Secondly, eachstage classifier in coarse detector mentioned above is taken as a feature, and a fine pedestriandetector based on those features is learnt with SVM from those training data which pass throughthe coarse pedestrian detector. Finally, color and spatial information based temporal analysisis utilized to improve the pedestrian detection rate and decrease the false alarm rate ulteriorly.Experimental results show this method high performance.3. Body Part Based Pedestrian Detection with Monocular Moving Camera A coarse-to-fine method for pedestrian detection is proposed in such environments. An in-dividual human is modeled as an assembly of natural body parts, including head-shoulder, torso,and leg. We introduce absolute Haar-like features and Edgelet features. Part detectors, based onthese features, are learnt from training images by Soft Cascade. Firstly, the pedestrian candidatesare generated by full-body detector. Then Bayesian decision based combination approach is uti-lized to determine pedestrians among those pedestrian candidates and can reduce the false alarmrate significantly. Experimental results show this method high performance in natural clutteredscenes.4. A 3D Feature-Based Tracking Algorithm Combining IMM and Cascade Data As-sociationA 3D feature-based binocular tracking algorithm is presented for tracking crowded peopleindoors. The algorithm consists of a two stage 3D feature points grouping method and a ro-bust 3D feature-based tracking method. The two stage 3D feature points grouping method canuse kernel-based ISODATA method to detect people accurately even though the part or almostfull occlusion occurs among people in surveillance area. The robust 3D feature-based trackingmethod combines interacting multiple model (IMM) method with a cascade multiple feature dataassociation method. The robust 3D feature-based tracking method not only manages the genera-tion and disappearance of a trajectory, but also can deal with the interaction of people and trackpeople maneuvering. Experimental results demonstrate the robustness and efficiency of the pro-posed framework. It is real-time and not sensitive to the variable frame to frame interval time.It also can deal with the occlusion of people and do well in those cases that people rotate andwriggle.5. A 3D Feature-Based Tracking Algorithm Combining IMM and MHTIf people are crowded within the surveillance region, they will interact each other, such asspace-closed people, occlusion or partial occlusion among those people. This will incur that adetection of person is assigned to the track of another person. Firstly, interacting multiple modelmethod (IMM) is adopted to predict the state of a person. Then multiple hypothesis trackingalgorithm (MHT) is adopted to match a person detection with the track of this person correctly.It also can generate a new track for a new entering person and delete the track of person thatgoes out of the surveillance region. The velocity continuity constrain and intensity consistencyconstrain are introduced to MHT algorithm to calculate the confidence of each track, whichcan reduce the number of hypothesis and thereby reduce the computational cost. Experimentalresults demonstrate the robustness and efficiency of the proposed framework to inter-occludedpeople and the nonlinear movement of people.6. MCMC and MHT Based Pedestrian Tracking with Monocular Moving CameraIn computer vision, the performance of pedestrian detection plays a very important role invariable number of automatic pedestrian tracking. However, pedestrian detector may bring a fewfalse negative and positive pedestrian detections since pedestrian has variety of shapes and ap-pearances. The pedestrian interacting and occlusion or partial occlusion between the pedestrianand scenes are also the challenging problems. A novel pedestrian tracking method combined the top-down approach and bottom-up approach is proposed, which can automatically initiateand terminate the track of a pedestrian if some conditions is satisfied. Firstly, this method trackspedestrian with top-down approach, which estimate the state of a pedestrian based on Kalmanparticle filter and Markov Chain Monte Carlo method; Then pedestrian is tracked with bottom-upapproach, which matches the results of pedestrian detection with the current estimate of existedtracks using multiple hypothesis tracking (MHT). If a track has not a corresponding pedestriandetection, this method uses the current tracked result by top-down approach as the state esti-mate of this trajectory, otherwise updating the current state of this pedestrian trajectory usingthe matching result by MHT. Therefore, this tracking method can overcome the false negativeand false alarm brought by the pedestrian detector. It also can correctly track pedestrians evenif there are pedestrian interacting and occlusion or partial occlusion between the pedestrians andscenes.