Automatic detection of anomalous behavioural events for advanced real-time video surveillance

This paper introduces the architecture and the implementation details, of an automatic real-time video surveillance system, capable of autonomously detecting anomalous behavioural events. Today video surveillance systems show two main drawbacks, namely: they are not flexible in adapting to different operative scenarios (they only work in a well known and structured world); they generally need the assistance of a human operator in order to recognize and to tag specific visual events. This paper describes an innovative approach for overcoming the previous limitations. In particular, the proposed system is capable of automatically adapting to different scenarios without any human intervention (but the placement of the TV sensors), and uses robust self-learning techniques to automatically learn the "typical" behaviour of the targets in each specific operative environment. Starting from this learned knowledge, the system can give alerts in an automatic way, by detecting the "anomalous trajectories" of visual targets in the controlled scene. To obtain robust self-learning capabilities, an improved version of the Altruistic Vector Quantization algorithm (AVQ) is proposed, capable of automatically describing the trajectories of moving objects in complex, not structured, outdoor environments. The modified AVQ is capable of autonomously calculating the number of trajectory prototypes, and improves the representativeness of the prototypes themselves. The final prototypes characterize the "typical" visual behavior of the targets in the controlled scene. Anomalous behavior is detected if visual trajectories deviate from the "typical" learned prototypes. The system has been implemented by means of standard PCs and TV cameras, and has been tested in many real outdoor contexts in different conditions (night and day). It is capable of running in real-time (15 fps for each camera) on general purpose HW. After a preliminary period (about 40 minutes in order to grant a significative interval of time to learn all the "typical" visual trajectories), the system is capable of giving automatic alerts about events that do not conform to typical behaviors. After changing the field of view, if enough learning time is granted (usually the previously mentioned 40 minutes), the sys- tem relearns the new scenario without any human intervention (no thresholds or other settings) and accurately detects anomalous events.