Object Detection And Tracking Algorithms For Smart Camera Networks

Smart camera is more appropriate for the next generation intelligent video surveillance system for its low-cost, convenient deployment and advantages of processing signal locally. Compared to object detection and tracking algorithms based on PC, the algorithm based on smart cameras has to face with challenges of limited memory and CPU computing capability. In this thesis, we take advantage of the s ignal processing characteristics of the embedded device, and design our object detection and tracking methods. The main work of this dissertation can be summarized as the following three aspects:1) The design of light-weight detection algorithm under the constrained resources of smart camera. An effic ient hierarchical light-weight background subtraction approach, called Block Compressed Sensing Background Subtraction, is proposed to balance the effic iency and the real-time requirement. The proposed algorithm combines the block-level and the pixel-level background subtraction modules into a single framework for extracting the object region efficiently. In order to save memory and improve computational effic iency, we introduce the block compressed sensing theory for the block-level module. Experimental results on various videos demonstrate superior performance of the proposed algorithm and only require several bytes per pixel.2) The design of object tracking algorithm when fac ing the challenge of limited resources and severe occlusion. In fact, we usually lack of enough effective information about the target, which means that we need an online learning strategy to achieve an effic ient tracker. Facing such challenges, we proposed a robust tracker based on online binary-SVM. Moreover, a simp le generative tracking model is embedded into a discriminative tracking model to form the coarse-to-fine searching strategy for improving the effic iency of the proposed algorithm. Experimental results on various videos show that the proposed tracker is robust to severe occlusion, pose variation and illumination changes.3) The installation of Smart Camera Network platform and the implementation of the light-weight object detection algorithm on it. We have applied several optimization strategies to improve the performance of the object detecting method during the DSP implementation, and meet the real-time requirement with 25 FPS.