Research On Computer Vision-Based Human Tracking System

In recent years video tracking of human which combines computer vision,artificial intelligence, pattern recognition and many other technologies becomes animportant research direction in the field of computer vision. It has great value in thearea of security monitoring and suspicious behavior determination. It also haspractical significance to safeguard national and public safety. Most of the currentproducts real-time monitor what is happening in the field based on high-speedautodome. It depends on people greatly and only can track the first target who goesinto the field automatically.The design needs to balance real-time and accuracy of the tracking algorithmsimultaneously, extract the effective features of the target accurately and make thesystem adapt to environmental changes. The above-mentioned several aspects aredifficult issues in design. This paper researched the first two aspects. Details are asfollows:1. Firstly this paper built system hardware platform and proposed systemimplementation plan which adopt different tracking strategies to the situation of singletarget tracking and tracking the specified target when multi-target existedrespectively.2. For single target tracking, the design adopted frame difference method to detectobjective, used corner point within the target area as the target feature and selectedoptical flow tracking based on feature to complete tracking.3. For specified target tracking when multi-target existed, the design allowed theuser to specify the target, simulated human head contour with vertical ellipse,extracted the gradient feature of the ellipse and the color feature inside the ellipse. Thedesign proposed different evaluation methods and measurement algorithm fordifferent features. The system used these above-mentioned features combined withparticle filter to complete tracking.4. In particle filter algorithm, we adopted particle's position and velocity in theimage as the state vector and established system model according to the state vector. Itis proved that adding velocity factor to the system model can receive better tracking results. We chose color feature inside the area of target ellipse as the observationmodel and used Bhattacharyya distance as measurement standard to describe thesimilaritybetween particles and target template. It provided basis for updating particleweights. This paper proposed a new method that redistributing weights according tothe weights size of remaining particles to resample particles. The experimental resultsshowed that this resampling strategycould improve tracking precision.5. Control the high-speed autodome rotate to maintain the target center in thefield of view according to the tracking results in the current frame.6. This paper conducted a lot of validation and comparative experiments. Theseexperiments provided a factual basis for tracking algorithm design, ensured thefeasibilityand robustness and verified the accuracy and real-time of the algorithm.The results demonstrated that the design of the human tracking system in thispaper can meet the tracking requirements when the target is moving, rotating andpartial occlusion in general environment.