DCT Based Fast Object Detection and Segmentation Design for Compressed Video and Implementation on Embedded System

• Main Authors: Hsun-Heng Tsao, 曹勳恆
• Other Authors: Ming-Hwa Sheu
• Format: Others
• Language: zh-TW
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/05453722123758727735

碩士 === 國立雲林科技大學 === 電子與資訊工程研究所 === 99 === Intelligent surveillance system has been gradually becoming an important role lately. It replaces the regular surveillance system which is monitored by manpower. As high technology keeps evolving, the surveillance system nowadays transfers real-time video by network instead of NTSC cable. In conventional way, surveillance center receives the network packets from remote network camera then being parsed into series compressed video data. After decoding the compressed video which is normally called as raw data of video, the algorithm of object detection and segmentation would be developed beneath it. The thesis proposed two novel methods for the object detection. At the beginning, we proposed our object detecting method with block-based texture at Spatial Domain, and then prove it is possible to translate into frequency domain for detection, with the advantage that the detection method of frequency domain analysis can implant into the MPEG Decoder at surveillance center. Our approach can achieve two times of processing speed than the conventional approach in spatial domain. Besides, it can also reach average 60% accuracy. The proposed approach is not only used for the dynamic background, but also implemented on embedded system “TI TMS320DM6446 DaVinci”. The performance on PC has exceeded real-time operation of 40 fps with frame size 768*576. Similarly, the performance on DaVinci embedded system can achieve 20 fps.