Research On The Technology Of Video-based Intelligent Scene Analysis And Human-Computer Interaction

With the explosive growth of video, people pay more and more attention to the intelligent video analyzing and processing. The automation of intelligent video analyzing and processing will dramatically decrease people's workload. We do research of key algorithms in two important regions of intelligent video analyzing and processing:video-based intelligent scene analysis and video-based human-computer interaction.After analyzing the video-based intelligent scene analysis (the software algorithm of intelligent video surveillance), we extract its key algorithms, and then do research on them:We present a foreground detection algorithm based on adaptive dual-threshold. It solves the problem of gaps and holes produced by many other foreground detection algorithms. The foreground point with high-confidence can be detected by the higher threshold. Meanwhile, the lower threshold and multi-level images will have the result that can connect the foreground points.In order to track multi-object robustly, our tracking algorithm adopts two level tracking method, which combines color-histogram based tracking and blob tracking. Firstly the blob tracking is used to do the first level tracking. Then the color-histogram is employed to distinguish the objects which can't be distinguished by the blob tracking.Employing our foreground detection and object tracking algorithm, we also develop three key functions in intelligent video surveillance:people-counting, illegal crossing line detection and removed object detection. Experiment results demonstrate that our algorithms are robust and practical, also meet the requirement of application.Moreover, we concern about the video-based human-computer interaction and use the posture in video as the input of interaction. We present an algorithm based on single camera. Dealing with complex background, we propose a multi-cue method which is based on velocity weighted features and color cue, to track the hand. Then we segment the hand using both Bayesian skin-color model and the hand tracking result. Finally, we use a novel method based on density distribution feature to recognize hand posture. It largely enforces the robustness of hand posture recognition because of scale-invariance and rotation-invariance. Experiment results demonstrate that our algorithm can deal with complex background particularly with large skin-color alike objects. Also Experiment results demonstrate our algorithm is real-time and robust.