| Summary: | 碩士 === 國立中央大學 === 電機工程學系 === 105 === Multi-object detection and occlusion tracking in the computer vision field is an important research topic, but most objects tracking algorithms are too complex and not practical for the real-time tracking system. This paper proposes a real-time occlusion-adaptive tracking method approach to resolving this issue. This method mainly improves the foreground detection to get low-complexity and high-quality effect. It also compares with other background subtraction techniques. Experimental figures show this method outperforms other foreground detection methods in terms of both computation speed and detection rate. For tracking moving objects, the proposed method uses the labeling to eliminate noises and group moving objects. In addition, it also proposed the processing cases of occlusions, including staggered case, separation case and multi-object in single label case, by using object's trajectory and edge. With this method, we can track the moving objects in the successive frame without color cues and appearance model in the real-time surveillance system. Finally, we implemented the hardware architecture of the foreground detection algorithm with a 150 MHz operating frequency at TSMC's 90 nm process. In this operating frequency, we can process 1080p image with 30 frames per second. And we are also working on the FPGA (Altera Sockit) for verifying. The results will be displayed on the screen through the VGA in 50MHz operating rate.
|
|---|
