Robust feature extraction and classification based automated human action recognition system for multiple datasets

Abstract

Action recognition schemes enable several intelligent machines to recognize human action by using daily life videos. In the last decades, recognizing human actions in the video sequences have been a challenging problem because of its real-world applications. Several action representation techniques have been done to improve action recognition performance. Approaches based on local features and classification are used for action representation, but it failed to capture temporal relationship between actions. In this work, Video Action Recognition (VAR) is assessed by using Weizmann, KTH and own datasets. Initially frames are extracted from input videos and the frames are resized. After the pre-processing, the object detection is done by Blob detection algorithm and tracking the object frame by frame is done using Kalman filter. The features are extracted from the moving object using feature extraction algorithms such as Bi-dimensional Empirical Mode Decomposition (BEMD), Scale Invariant Feature Transform (SIFT) and Discrete Wavelet Transform (DWT). These feature extraction techniques were applied on pre-processed frames to extract the efficient features from multi-scale images. Finally, the features are given to Convolution Neural Network (CNN) classifier for action recognition prediction. The proposed method is called as Hybrid Feature Extraction (HFE) and CNN used for VAR (HFE-CNN-VAR) method. The experimental results showed that the HFECNN-VAR method improved the accuracy in action classification. The classification accuracy is 99.01% for KTH dataset, 99.33% for Weizmann dataset and 90% for own dataset.