Visual Perception Based Human Behavior Understanding

Video surveillance systems are receiving wide attention as one of the most common security and protection measures with the continuous development of network and computer related technologies and the increasing demand of public security.The development of intelligent surveillance has been gotten more and more attentions.It is very pivotal to recognize some abnormal behaviors from surveillance videos in visual-based behavior understanding technology.Some related references at home and abroad have been read and investigated extensively.Visual perception based behavior understanding of moving object is studied in depth aiming at some drawbacks and limitations for abnormal behavior recognition in the video surveillance at present.Some main features and contributions of this dissertation are as follows.A feature extraction method has been proposed based on fluid dynamics Lagrangian dynamic particle flow.The motion of the foreground object from the video is mapped to some dynamic particle flows that effectively reflect the change state of its motion after both optical flow and dynamic particles being combined.The motion state of the moving object is reflected by analyzing the motion trajectory of particles.A novel strategy for placing and deleting particle dynamically has been developed to adjust adaptively the number and the density of particles in the video scene to improve the computational efficiency.The experimental results show that the proposed method can be applied to extract the motion behavior features effectively on the premise of independent recognition of moving objects and motion tracking with high efficiency and performance.Some recognition requirements of abnormal movement behavior can be met based on common hardware platform.An on-line detection method has been developed based on video of sparse particle flow field.Some motion information is analyzed statistically from the bottomed sparse particle flow field.Some unstable motion features of the reactive moving objects are extracted along with some background interference information being removed.The dynamic threshold strategy has been proposed to distinguish the positive behavior from some negative ones.Some adverse effects are overcome including ambiguity and scene diversity at the actual monitoring environment.Experimental results show that the proposed method is effective and feasible.A dynamic Bayesian decision based method for video anomaly recognition and localization has been proposed.It is difficult to describe the features of manual work due to the fact that the features of single motion cannot effectively reflect the complete state of motion and the ambiguity of abnormal behaviors.The basic element characteristics of the reaction state are extracted from the optical flow field.Multi-channel motion features are constructed based on motion direction,motion amplitude and motion resistance.Gaussian mixture model based multi-channel motion features are learned on-line.A statistical model of motion features based on image sub-region is obtained.The abnormal behavior recognition problem is transformed into a Bayesian decision one.The motion characteristics are determined as abnormal behaviors that deviate from the above statistical model in the sub-region of the image.Experimental results have been shown that the proposed method has excellent identification performances.A time coding method for motion trajectory has been developed to integrate both spatial and domain change information of particle motion.The motion trajectory is employed to describe how the object moves and changes its position in space,which reflects the motion pattern and energy spatial distribution.The motion time coding is performed based on the motion trajectory of dynamic particles.The motion evolution model has been presented in a form of visual behavior template.The visual features are enhanced by pseudo-color coding.The proposed time-coded trajectory method can improve effectively the recognition effect compared with space motion trajectories.A double-flow convolution deep learning based anomaly behavior recognition method has been proposed.The Lagrangian particle trajectory is combined with time series dynamic evolution.The time evolution pattern of movement behavior in video scene and its dynamic trajectory evolution diagram are obtained based on color coding strategy as mentioned above.The video features are gotten by smoothing the time-varying mean vector data.The convolution pooling strategy is adopted to integrate the apparent feature information and the motion feature one of the moving object after the double-flow framework of deep learning being combined.It can be perform to realize automatic recognition of motion anomaly in video monitoring scene.Some experimental results have been shown that the developed method has robust motor behavior pattern description and excellent identification capability for abnormal behaviors through experimental comparisons and analyses.