Action Recognition Method Based On Sparse Auto-Combination Spatio-Temporal Convolutional Neural Network And Its MapReduce Implementation

Automatic human action recognition system has a very important practical value in many applications. However, the current action recognition methods still possess shortcomings of this kind or that kind. Studying how to develop a robust action recognition method is still a open problem and has a significance of promoting machine learning. Convolution neural network (CNN) is a deep learning model, it is inspired by vertebrate visual nerve system, it can learn advanced abstraction features from raw images and has a powerful images classification capacity. However, it is limited to classification of single image. In order to apply CNN to action recognition of video, we added some improvements to CNN. The contribution of the thesis are as following several aspects:We proposed a spatio-temporal convolutional descriptor which can extract complexity advanced abstraction features from video and developed a spatio-temporal convolution neural network (STCNN) based on spatio-temporal convolutional descriptor for action recognition. Experimental results showed the effective of STCNN. In order to enhance STCNN, we further proposed a sparse auto-combination strategy to convolve multi input maps in the stage of convolution. Using the strategy that imposing a kind of sparsity limitation in convolving multi input maps, so the convolution layer has the ability to learn the optimal input maps combination as input, and save the trouble of complex manul setup steps comparing to conventional strategy. Experimental results showed that the sparse auto-combination spatio-temporal convolutional neural network (SASTCNN) is superior to STCNN.In the age of big data, for the purpose of dealing with large-scale of video data, we further proposed a parallel matrix multiplication algorithm with MapReduce (MMMR) and implemented SASTCNN on Hadoop platform based on the MMMR algorithm, we called it SASTCNN-MR (SASTCNN with MapReduce). We comparing the experimental results with SASTCNN and found that SASTCNN-MR is accurate, stable and efficient. In order to take advantage of the high computing power of CPU with multi core, we used multi-thread to implement MapReduce process and used the new MapReduce API to test SASTCNN, we called it SASTCNN-MRMC (SASTCNN-MR with Multi Core) and the experimental results showed a better performance.We experimented on two dataset-WEIZMANN and KTH which is stand dataset for action recognition. The experimental results in different scenarios was displayed. We arrived at a conclusion that the proposed method is competitive to other based line method and more flexible.