Research On Action Recognition Based On 3D Depth Video Sequences

In recent years,human activity recognition becomes a hot research direction.More and more researchers become to devote themselves in this area.However,human activity recognition is not a simple task.The diversity of human activity and the influence of environment factors such as sunlight and shade increase the difficulty.With the arrival of the 21 st century,the level of technology and the production capacity have made a huge leap,especially dramatic changes in computer technology,sensors and multimedia.The hardware devices commonly used in action recognition,such as RGB cameras,depth cameras,range sensors and wearable acceleration sensors,make the technology of action recognition become reality gradually.Due to the superiority of the depth cameras,in this paper we focus on the algorithms of human activity recognition based on depth data.The main innovation points of this paper are as follows:(1)Since the traditional depth motion maps(DMMs)compress the whole movement sequence into a two-dimensional image,the time information is mostly lost.So this paper proposes a new feature extraction algorithm based on vague division.First,sequences are divided to extract depth motion maps.At the same time,with the inspiration from HOG which uses an overlap way to prevent a continuous region from being divided separately,our algorithm proposes a vague division strategy for the video sequence.Thus,there is no specific boundary between divisions and frames can be shared between them.This makes the algorithm can store continuous time information and the extracted feature will be more convincing.(2)In order to solve the problem of energy inhomogeneity between vague divisions,this paper proposes a new division method utilizing motion energy based on the vague division strategy.Different from the first work which divides the sequences based on the specific number of frames,our second work makes divisions considering the energy between frames and the energy is more evenly distributed.Therefore,the resulting vague boundary sequences have strong adaptive ability and can be automatically divided according to the energy distribution.In this paper,we denote these vague boundary sequences as self-adaptive vague boundary subsequences.(3)Although the algorithm based on the vague division strategy achieves a good performance,the details including the motion speed can be lost upon such a single scale division strategy,especially the complex motions.In order to solve this problem,this dissertation proposes a human activity recognition method based on self-adaptive vague boundary subsequences.This algorithm first extracts the self-adaptive vague boundary sub-sequences and then uses the multi-temporal DMMs to capture more details.LBP features and Fisher vectors are used to describe the depth motion maps to obtain the final feature vectors.To deal with the small sample problem in human activity recognition,a robust classifier named probabilistic collaborative representation classifier(R-ProCRC)is finally used.The results of the experiments proved that the proposed method based on self-adaptive vague division multitemporal DMMs can acquire a satisfying recognition result on several common public datasets.