Still Image-based Human Activity Recognition with Deep Representations and Residual Learning

Iterative Recognizing human activity in a scene is still a challenging and an important research area in the field of computer vision due to its various possible implementations on many fields including autonomous driving, bio medical, machine intelligent vision etc. Recently deep learning techniques have emerged and successfully deployed models for image recognition and classification, object detection, and speech recognition. Due to promising results the state of art deep learning techniques have replaced the traditional techniques. In this paper, a novel method is presented for human activity recognition based on pre-trained Convolutional Neural Network (CNN) model utilized as feature extractor and deep representations are followed by Support Vector Machine (SVM) classifier for action recognition. It has been observed that previously learnt CNN knowledge from large scale data-set could be transferred to activity recognition task with limited training data. The proposed method is evaluated on publicly available stanford40 human action dataset, which includes 40 classes of actions and 9532 images. The comparative experiment results show that proposed method achieves better performance over conventional methods in term of accuracy and computational power.