A Research Of Action Recognition Algorithm Based On Optical Flow Guided Feature

Behavior recognition is an important research topic in the field of artificial intelligence and computer vision.In daily life and work,smart devices with behavior recognition capabilities play an important role in scenarios such as video retrieval,video surveillance,and human-computer interaction.Existing behavior recognition algorithms generally have the problem of slow running speed,which cannot achieve satisfactory realtime results in practical applications.In response to this problem,this paper proposes a behavior recognition algorithm based on optical flow-oriented features.The main work is as follows:Aiming at the problem that the speed of extracting optical flow features is slow and it is difficult to achieve a satisfactory real-time effect,this paper uses optical flow guided features(Optical Flow guided Feature,OFF)as a medium to represent motion information,and proposes an OFF feature differential extraction algorithm.The algorithm first uses the Sobel operator to extract spatial features,and then obtains the timing information between video frames through differential operations.Finally,the spatial features and temporal features are fused to obtain the required OFF features.In this paper,a sparse sampling strategy and the proposed algorithm are used to construct an Action Recognition Network based on differential extraction algorithm(ARNDEA network).Through experiments,the position,learning parameters and number of iterations of the OFF feature sub-network in the network are optimized,and the fusion method of spatial and temporal features in the network and the accumulation method of the OFF feature sub-network are explored.Experimental results show that the optimized network gets better accuracy and speed on the test data set.According to the built ARNDEA network,this paper implements an end-to-end behavior recognition system.The system only needs to input the sample to be tested into the trained network,and the action label of the corresponding sample can be quickly obtained,and the real-time processing effect can be achieved.Finally,this paper has carried out experimental verification on the proposed ARNDEA network,which has achieved 95.3% and 70.2% accuracy rates on the UCF-101 and HMDB-51 data sets,respectively,and a recognition speed of around 120 fps.In summary,the method proposed in this paper has achieved good results in terms of the effectiveness and real-time performance of behavior recognition.