Abstract:

Over the last decade, there has been a fast growth of surveillance webcams in every part of human activity, resulting in a massive increase of camera footage. The main objective of this paper is to give an outline and comparative evaluation of new Deep Learning techniques for Human Activity Recognition (HAR) for different types of datasets. Our Novelty lies in the Evaluation and comparison of different Deep Neural Networks. HAR has several applications like Human Computer Interaction (HCI), Smart Driver Assistance Systems, Personal Assistant, Interactive games, Content based Video Annotation, Smart Medical Assistance systems, Smart Office systems, Smart Traffic Monitoring systems, Sports Analysis and Crime Scene Analysis. Following that, we examine and categorize current options offered to meet these objectives. All these applications face common set of challenges like noise, illumination changes, multi-view camera angles, partial occlusions, semantic classification of human activities. All the works are mostly based on Spatial and Temporal feature data for Activity recognition. Extraction of both temporal and spatial data from surveillance videos is required for successful video categorization. As a result, this work (CNN, RNN, LSTM, LSTM, and Multi-Stream Convnet Architecture) examines recent improvements in hierarchical conscience content deep learning architectures. It also dives into the various deep learning models available for HAR. We also go through the datasets (KTH, HMDB-51 and UCF-101), that were used for evaluation. This study seeks to provide the groundwork for future media HAR by identifying important challenges in the efficacy of human event detection in image sequences utilizing deep learning models. Finally, we present the accuracy comparison of different state of art techniques in HAR, followed by the disadvantages of current methodologies in the Deep learning approaches and conclude with futuristic trends in the human activity recognition.

Description:

CNN, 3-D CNN, Deep Learning, Human Activity Recognition, LSTM, Multi-Stream Convnet, RNN, 2-Stream Convnet