Alogrithm and Architecture of Video Segmentation Using Model-Based Sprite Generation for Advanced Surveillance Systems

Video surveillance systems can help people to monitor environment and provide people more security of life. In advanced surveillance systems, content-based video processing techniques must be added to equip the surveillance systems with intelligent functions.

The goals of advanced surveillance system are to become automatic and large in scale. To become automatic, video segmentation should be included into surveillance systems. The result object masks can be used as input for object tracking and behavior analysis algorithms. To become large in scale, some computations should be distributed to cameras to reduce computational load on the server. Since video segmentation is necessary and independent of application situations, we think that it is very suitable to be integrated into cameras.

In this thesis, the algorithm of video segmentation using model-based sprite generation is proposed. This algorithm targets at surveillance cameras, especially panning cameras. We utilize the application characteristics to simplify the sprite generation process. That is, a surveillance camera usually has fixed position and known moving pattern, which are both considered in the proposed algorithm. For cameras with fixed position and panning motion, the captured frame can be first mapped to cylindrical coordinate by cylindrical warping. After cylindrical warping, the camera motion model of global motion estimation (GME) operation can be reduced from four or six parameters to only one parameter. With sprite as the background for segmentation, several object detection problems also occur, and we also proposed an adaptive threshold and sprite updating scheme. Adaptive threshold adjusts threshold value according to the current frame situation, where sprite updating scheme can constantly update sprite to generate better segmentation results.

Corresponding architecture for video segmentation using model-based sprite generation is also proposed. Since this algorithm is designed to be integrated into cameras, the main hardware design goal is low cost. Mixed pipeline scheme is
used to reduce buffer requirement for the whole system. In addition, hardware architecture design of model-based GME is also proposed, where we proposed slice buffer to solve the buffer wasting problem due to irregular memory access.
The model-based GME is implemented with TSMC 0.18 μm 1P6M technology. The chip size is 2.08×2.08 mm2. This chip is capable of processing 120 QVGA frames or 30 VGA frames in one second.