Davinci Platform Based Video Anomaly Event Detection Algorithm Research And Implementation

The wide application of video surveillance cameras and the development of Intelligent Video Surveillance Technology(IVS) led to video surveillance market booming. Video surveillance monitor systems, which take people as the main monitor body, are no longer have the ability to handle real-time processing of mass surveillance videos inputted from hundreds of thousands all-weather surveillance cameras. As an important branch of intelligent video surveillance, video anomaly event detection can make use of computer vision technology to detect the small amount of anomaly behavior events that do not match normal behavior events automatically, and send out alarm signals in time, thus freed people from sitting in front of the screen monitoring of boring work.The main concrete works are as follows:1.By analyzing anomaly event detection algorithm theory and application advantages and disadvantages based on fixed-location monitors, which proposed by A.Adam, we found that in different surveillance environments, equidistant arrangement of monitors may result in information loss and calculation redundancy. Based on it, we proposed a self-organization plan based on surveillance environment, which make automatically adjust the position and density of the observation points come true.2. Based on SEED-DVS6446 DaVinci board, the anomaly event detection algorithm based on observation point developed which can runs on its DSP side, and anomaly detection system developed which can run on its ARM side. At last, we get a video anomaly event detection box, which can real-time detect the video streaming whether has abnormal events happens or not after turning on its power, and anomaly area will also be figure out once anomaly event real detected.3. We Used Gaussian Mixture Model for extracting foreground moving blobs and optical flow for calculating the move direction of them. "Hog + linear SVM" and "Haar + Cascaded AdaBoost" programs are used for pedestrian and vehicle detect on moving blobs’ image respectively. Through blob tracking for detected pedestrian or vehicle, we obtained its history trajectory in video scene. Combined with the anomaly event behavior rule set which we summarized, we achieved such describable anomaly event distinguish from video streaming like pedestrian and vehicle in break, pedestrian and vehicle mix line, pedestrian linger in surveillance environment.From theory to practice, by integrating the former three parts, we finally achieve a complete video anomaly event management system, which links observation point algorithm based anomaly event detect system runs on SEED-DVS6446 DaVinci board, and specific describable anomaly event discrimination based on moving object detect and track together, and runs in concurrently. Experimental performances confirm the system can be used in video scene for real-time detection of both known and unknown anomaly events.