A Comparison of Two Neural Network Based Methods for Human Activity Recognition

In this paper, two different methods are introduced for human activity recognition based on video signals. Method 1 explores the effectiveness of combining feature descriptors obtained by local descriptors and artificial neural network classifier. It stays in traditional approach that is local descriptors extract interest points or local patches from videos, then feature vectors are constructed based on them, and eventually feature vectors are used as the input of a two-layer feed-forward artificial neural network (ANN). Experimental results show that using HOG3D descriptor with ANN gives the best performance. On the other hand, deep learning architectures have attracted much consideration in the last years for automatic feature extraction, so an improved 3D convolutional neural network architecture is also designed as method 2. They are implemented and compared with state-of-the-art approaches on two data sets. The results exhibit that method 1 is superior when the shortage of sample data is the main restriction. It achieves recognition accuracies of 97.8% and 99.8% for the Weizmann and KTH action data sets, respectively. Also method 2 is considerable because of its automatic features extraction and achieves an acceptable result for video with lots of original training data. So that it gets recognition accuracy of 92% for the KTH data set while this value is drastically reduced for the Weizmann data set.