| With the increasement of public places and human activites,intelligent surveillance system is increasingly becoming an important part of securityin public places. Most of the current intelligent surveillance systems are used formonitoring pedestrians, including pedestrian detection, pedestrian tracking, abnormalmotion detection and so on. Pedestrian tracking and abnormal motion detection invideo surveillance are hotspots in the field of computer vision, and attract more andmore attention.This paper has a deep research on pedestrian tracking and abnormal motiondetection combined with the latest computer vision and machine learning researchtheory. We proposed corresponding improvements and solutions based on theanalysis of existing algorithms of pedestrian movement analysis.(1) In the aspect of pedestrian detection, we make an improvement in classicpedestrian detection method based on Histograms of Oriented Gradient (HOG). Thecandidate region of pedestrian is detected according to the symmetry of human body.The gradient feature of candidate region is extracted by HOG, and then used by linearSVM classifier for pedestrian detection.(2) In the aspect of pedestrian tracking, We propose to formulatemulti-pedestrian tracking as an minimization of energy function. Other than anumber of recent approaches, we focus on designing an energy function thatrepresents the problem as faithfully as possible. The energy function can representthe motion and interaction of all objects of interest in the scene. The optimaltrajectories can be found by optimizing the energy function iteratively.(3) In the aspect of abnormal motion detection, we apply the probabilisticgraphical model to abnormal motion detection. Probabilistic graphical model candescribe the dependencies of random variable abstractly in diagram and can modelstate changes of pedestrian movement well. The HMM and HCRF can be trainedwith the features vectors of the motion, and at the last we compare the discriminationeffect of both models. |
PDF Full Text Request