Research On Human Tracking And Action Recognition Based On Video Sequence

The object detection in video is an active research area in the field of computer vision in recent years,the action recognition algorithm,which realizes the understanding of video content,has more and more wider application space in scenes such as human-computer interaction,video annotation and abnormal behavior detection.With the continuous deepening of the theoretical study of pattern recognition and deep learning,tracking and recognition algorithms are also innovating and gradually developing into practical applications.Despite this,there are still some difficulties to be solved,such as mining more powerful feature expressions and dealing with challenges such as occlusion and scale transformation.Although the neural network algorithm is widely used in the current field,it still has a lot of room for improvement in the speed of operation and the need of input sample amount.Based on these considerations,this paper attempts to improve on the basis of existing object tracking and behavior recognition algorithms.For the problem of convolution features on the lack of image characterization capability,this paper implements a self-adaptive fusion algorithm of depth features and LBP texture features based on correlation filtering templates.The algorithm uses different convolutional layers in the convolutional neural network structure to characterize the target area.The shallow features describe the location information,and the deep features contain more semantic information.Considering that the depth feature can not solve the problem of target deformation,the LBP feature describing the local texture feature is introduced to make up for this deficiency.Finally,an adaptive fusion method is adopted to fuse the two features at the decision-making level of the algorithm,which is better than additive or multiplicative fusion.In order to solve the problem that there are noise points in dense trajectories which affect the recognition accuracy,the dense trajectory points are purified by motion saliency value to remove points in background area and the motion disturbance area in an image,and a two-stream neural network is trained based on the dense trajectories after purification,which can make full use of both spatial and temporal information from the image sequence.In this way,a more comprehensive video behavior representation feature is extracted,and the features are eventually input into a support vector machine classifier,which uses an OvR strategy to obtain the multiple classification results.Through experimental verification,the object tracking algorithm in this paper has obtained improvement of multiple indicators on the OTB-50 dataset.The CLE is reduced by 52.2 pixels compared with the LBP feature method,50.1 pixels compared with the DLT algorithm,the OS is 36.9% higher than the LBP feature method,28.4% higher than the DLT algorithm.It is also more robust when encounters illumination variation,scale variation,occlusion,in-plane rotation,and object deformation.The action recognition algorithm also achieved higher classification rate on HMDB and UCF101 datasets,which are 5.4% and 3% respectively compared with the basic two-stream neural network.