A method of motion recognition and its parallelization in sparse self-assembled spatiotemporal convolution neural networks

A method of motion recognition and its parallelization in sparse self-assembled spatiotemporal convolution neural networks

Sanding of Xiamen University

In order to improve the performance of spatio-temporal convolution neural network, a sparse self combination strategy is proposed when the input feature graph of the convolution layer is combined. By adding sparsity restrictions to the input feature graph, the convolution layer can automatically learn the best feature graph combination as input, compared with the traditional manual setting method, the complicated steps of manual setting are omitted, the experiment shows that the spatiotemporal convolution neural network with sparse self combination strategy has better characteristic learning ability and classification ability. This paper presents a matrix parallel multiplication algorithm based on MapReduce, and based on the parallel algorithm, the MapReduce programming model is used to parallel the sparse self assembled spatiotemporal convolution neural network in the Hadoop platform, and compared with the serial experimental results, The feasibility, stability and correctness of the parallelization of the sparse self assembled spatiotemporal convolution neural network are validated, and some speedup ratios are obtained. In order to utilize the computing power of multi-core CPU, the MapReduce map process and reduce process are implemented by multithreading, and the algorithm is applied to the training test of sparse self assembled spatiotemporal convolution neural network, and the performance is further improved. A series of experiments were carried out on two public datasets of Weizman and KTH, showing the performance of spatiotemporal convolution neural networks in various scenarios. Experimental results show that compared with other datum methods, the method presented in this paper shows very competitive results in two datasets.

A method of motion recognition and its parallelization in sparse self-assembled spatiotemporal convolution neural networks