Event Detection Algorithm For Surveillance Video

With the rapid development of city economy, the enlargement of city scale and the increasingly growing of city population, human-based traditional surveillance method has become difficult to meet the needs of the times. Surveillance video based event detection system can analyze and interpret video content automatically and detect abnormal events in time, thus making up the shortage of human manage resources and solving some drawbacks of traditional surveillance methods. Due to the great learning significance and practical value of event detection algorithm, we mainly study five typical kinds of event detection, crowd counting, pedestrian detection, fall detection, object movement detection and limitline acrossing detection in this thesis. The main tasks are as follows:1. We propose an improved method of crowd counting based on features and regression analysis. A new low level feature, the number of corners is incorporated to highlight the correspondence between features and pedestrian number. RVR is introduced to regress the number and the estimated result is fused with that of GPR. Experimental results demonstrate our work outperforms the state-of-the-art methods.2. We have implemented and improved an APCF and cascade Adaboost based pedestrian detection method. APCF, a kind of Haar-like feature, can reflect the difference between object and background. Due to its weakness in describing edge information, we improved this algorithm by introducing Edgelet feature. Experimental results validate our improvement is effective and indicate that multi-feature fusion may become the main trend of pedestrian detection in future. 3. In this thesis, we propose a context-based fall detection system by analyzing human motion and posture using HMM and RVM respectively. Additionally, we integrate homography to deal with falls in any direction. The system is validated on an open fall database and our own video dataset. Experimental results demonstrate that our method achieves high robustness and accuracy in detecting different kinds of falls and runs at a real-time speed.4. We improve a kind of limitline acrossing detection algorithm. An object tracking module is used to locate and track interested object. The motion trajectory of object is recorded in every frame and acrossing event is detected by analyzing the relation between object trajectory and the location of limitline.5. We improve an object movement detection method by handling it as an object template matching problem. A HSV space color histogram and a rotation-invariant edge orientation histogram are extracted to build the appearance model for monitoring objects. The similarity between two objects is measured by histogram intersection. If the similarity is below a threshold for a certain period, we judge it as an object movement event and raise alarms.