Event Detection Algorithm For Surveillance Video

Home and public security have become a hot topic in recent years.Intelligent video surveillance system is an important part of the field.In this thesis,we mainly focus on three aspects:indoor people number statistics,fall detection and event detection in complex scene.1.The base of indoor people number statistics is pedestrian detection.Fast-DPM algorithm is used to detect pedestrians.To improve speed,we remove the background with background model.Then,we proposed a simple and fast algorithm to track the people who are not detected.Finally,we get the number of people based on the result of detection and tracking.2.We propose an important feature which called Overlapping Ratio based on Homograghy(ORH)to detect falls in surveillance videos.This feature can reflect both the fall speed toward the floor and the distance between fallen person and the floor.Posture feature is also extracted to remove false alarms.The experimental results demonstrate that our method achieves high accuracy in detecting different kinds of falls and runs at a real-time speed.3.The event detection in complex scene is very important in the field of video surveillance.In this paper,we detect events with trajectory analysis.In order to obtain better trajectory,we adopt the linear regression for the scene correction and adopt the Gaussian process regression to smooth the trajectory.For running events,we extracted the speed characteristics to detect events.For the events of "People Meet"and "People Split Up",we extract the observation features(such as speed,distance)and use the HMM model to model the process of the events.