Adaptive Vector Quantization of Sequences of Local Blocks for Video Surveillance System

To keep security, video surveillance cameras have been installed in public space such as convenience stores, streets, A TM , etc. The number of such cameras is increasing rapidly. W hen an incident happens, the circumstances of incidents can be confi rmed by review ing the recorded scene. H ow ever, many of the current surveillance systems record continuously everything w ith low frame rate (for ex ample 4 frames/ sec.) even w hen nothing happens. B ecause of lack of retrieval function, the current user of the surveillance system has to review every bite of information to fi nd the parts w here the incident is recorded. S ince a static camera is used for video surveillance in most cases, a lot of the information recorded is redundant such as the backg round of the scene and w hen nobody is in the scene. The review ing of such redundant information can create a loss of time w ith dramatic conseq uences. A lso the system often fails to record very important information such as frontal face because of its low frame rate. To overcome such draw backs of the current surveillance system, w e should develop a method to retrieve scenes w here some moving objects are recorded. A lso an effi cient video compression alg orithm has to be developed to increase its frame rate and reduce the req uired storag e. F or a case of static camera, a typical approach to detect moving objects is backg round subtraction. If there are some moving objects in the current imag e, the diff erences at the points corresponding to the moving objects become larg e. The moving objects can be easily detected by thresholding the diff erences. In realistic situation, how ever, the backg round may g radually or suddenly chang e because of the day lig ht chang es or movements of the eq uipments, etc. To treat such chang es in the backg round, the adaptive backg round estimation method has been proposed by one of authors [1 , 2 ]. In [3 ] a method is proposed to use G aussian M ix ture M odels w ith pix el values to modelize the backg round. In this paper w e address the problem of recording and retrieving scenes of interest from a static camera for a video surveillance system. H ere w e consider seq uences of local blocks instead of seq uences of pix el values in the case of backg round subtraction. To model a seq uence of a local block, w e use an adaptive vector q uantizer (A V Q ). B y applying a vector q uantizer to a seq uence of a local block, w e can represent the seq uence by some of code vectors (codebook). This corresponds to modeling multiple backg rounds appearing in the seq uence. Then novelty of a new block can be easily detected by checking the distances betw een the new block and the code vectors. This makes moving object detection possible. A lso by representing the seq uence using a codebook the req uired storag e is reduced and approx imation of all blocks in the seq uence can be reg enerated from the stored codebook. To maintain information on the current backg round, the code vectors are stored in a stack in the order of the occurrences. A lso a small memory of its history is assig ned to each code vector of the codebook allow ing a forg etting factor of the vector w hen it has not been used for a long time. W e also present a method to remove noise from the scene such as fl ickering by monitoring the codebook evolution throug h time.