Research On Human Behavior Recognition Combining Global And Local Features

Video-based human behavior recognition is one of the hot areas and important branches of computer vision research.It aims to allow computers to identify the movements and behaviors of human beings in video and it is widely used in many fields such as intelligent monitoring,video retrieval and virtual reality.With the continuous expansion of marketing demand and the improvement of computer computing capabilities,human behavior recognition has attracted many domestic and foreign scholars who have achieved unprecedented development.However,due to the complexity and uncertainty of human motion patterns and the variability of environment,the research on human behavior recognition faces challenges.In-depth research on feature extraction and feature encoding are conducted.Moreover,two complementary features,the local and global features of video,are fused to complete the recognition.Initially,the classical framework of human behavior recognition is studied and the current classic feature extraction methods are discussed from the perspective of global features and local features.Since the motion history image can well reflect the temporal and spatial position of human motion,it can also visually indicates the sequence of human motion.Therefore,the motion history image is used as a global feature for human behavior recognition.It is a static time template that is formed based on pixel changes between adjacent frames and is represented by its pixel brightness.Secondly,the dense trajectory algorithm taking into account the gradient histogram,optical flow histogram and motion boundary descriptor in each video space,and thus the spatial structure information of the video is well preserved.Therefore,the dense trajectory feature is used as a local feature to recognize the human behavior.The image pyramids with different scales are established and the pixels at each scale are sampled uniformly at regular intervals with a fixed step.Then,Shi-Tomasi criterion,one way to filter points,is used.In order to avoid the fail of tracking,the number of tracked frames is limited and different descriptors are derived based on dense trajectories.Furthermore,feature encoding is performed on all the features.Finally,the recognition of human behavior is studied based on dense trajectory feature and motion history image and series-level fusion method is adopted to form a new feature.It can be seen from the experimental results that human behavior recognition based on fusing dense trajectories and motion history image can achieve better recognition results.