Researches On Key Techniques Of Human Action Recognition In Videos

Human action recognition in videos is an important research branch in the field of computer vision and pattern recognition.It will be widely applied in many fields such as intelligent video surveillance,human computer interaction,physical training,and video retrieval.And it has attracted wide attention of researchers at home and abroad.However,due to the complexity and diversity of human body movement,human action recognition becomes a challenge research topic.We consider the human action videos as two categories as static videos and moving videos,according to whether the camera motion occurring during human action videos capturing.For static videos,the global representation method based on the body silhouette extracted by moving foreground detection is considered.While,for moving videos,the local representation method with video local features is considered.The main research contents and contributions are as following:1)Foreground detection based on spatio-temporal conditional information(SCI)in static videosTo detect the body silhouette in static videos,the moving foreground detection algorithm is researched.To handle with the dynamic background in video surveillance,a foreground detection method based on spatio-temporal conditional information(SCI)is proposed.The image pixel spatio-temporal conditional information is used to instead of pixel intensity for pixel classification.During SCI calculating,the visual salient principle is adopted to construct the spatio-temporal domain to increase the discriminate between background and foreground.And the neighborhood pixel weighted SCI(NWSCI)is used for reduce the noise in foreground.While an accelerate method based image block detection is adopted.Finally,it achieves real-time and high accuracy foreground detection in dynamic background videos.2)Global action representation based on Riemann manifold in static videosThe human silhouette space time shape(STS)is adopted for human action global representation in static videos.To handle with the high dimension problem in STS matching,a novel dimension reduction method is considered.Firstly,the local samples of space time shapes(Ls STS)with point light radiation distance in STS are extracted.Secondly,a covariance matrix of Ls STS is used to reduce the dimension of STS.Then human action classification is implemented on Riemann manifold by STS matching.It can achieve high accuracy human action recognition in static videos with low complexity.3)Camera motion invariant local video feature extraction and descriptionThe problem that current spatio-temporal local feature detection methods will detect the moving background pixels and generate many local features not related with human motion in moving videos is considered.To handle with these problems,a camera motion invariant video local feature(CMI-VLF)is proposed.CMI-VLF firstly detects image local features and track these features to generate feature trajectory,then classifies these feature trajectory as background or foreground by the subspace reconstructed method based on RANSAC(RANdom Sample Consensus).The foreground feature trajectory is related with human motion and used for CMI-VLF spatial location.The background feature trajectory is used for camera motion matrix estimation and image motion compensation.The descriptor of CMI-VLF is extracted from the compensated images.CMI-VLF can effectively increase the correct detection ratio of local video features and improve the performance of Bo F(bag of feature)for human action recognition in moving videos.4)Local feature space time coding methodIn order to solve the limitation of Bo F that it ignores the space time relationship of local features,a space time coding(STC)method for video local features coding is proposed.It involves space time locations of local features into feature coding to directly model space time relationship.The experimental results show that the proposed algorithm can effectively improve human action recognition accuracy than current feature coding methods such as vector quantization(VQ),sparse coding(SC),and locality-constrained linear coding(LLC).5)Human action classification based on video segmentation sequence setIn order to improve human action classification accuracy,a video segment sequences set based action classification framework is proposed.It divides an input video as a segment sequence set with overlap sampling,then implement action classification on this segment sequence set.It can increase the test samples and retain more human motion information for human action classification.Under this framework,the KNN voting classifier and locality constrained group sparse representation classification(LGSRC)model are adopted to improve human action classification accuracy.