Research On Violence Detection Algorithm Based On Trajectory Analysis

Human Action Recognition is an important branch in the field of Pattern Recognition. It is a combination of Computer Vision and Machine Learning. The goal is to structure the video stream, and analysis events in the clip using methods like machine learning.Violence Detection, in particular, is more difficult, due to the complexity of the targets and the motion pattern. There lacks of a generalused, high-efficient way to detect the violence action. To build a real-time violence detection system is important, for it can cut off labor costs, increase performance and shorten the reaction time, especially when real-time surveillance camera becomes regular.The paper provides a new violence analysis algorithm. First, it extracts the key-points of moving target, using ORB algorithm. With the FLANN algorithm, a great amount of motion trajectories are gained, which are the features to distinct violence behavior from non-violence ones. Then, we optimize the violence behavior’s trajectories, using short-trajectory filter and multi-paragraph least square method to improve the performance further. After that, we describe the trajectories with feature vectors of BRIEF descriptor and Markov chain model in both out-look and geography layer, and build a bag-of-word model. Finally, we put the vectors into SVM to train and classify. We adopt the muti-channel SVM to make different kind of vectors have different influence on the result, which leads to a promotion in performance.We experience it on the violence video dataset, HockeyFights, and gain a state-of-art performance.