Abstract:
Video analytics has become a vital field due to availability of the enormous video data. Automating human activity recognition from video footage is becoming more popular attributed to its numerous applications in surveillance, healthcare, and industries. The availability of huge amounts of data facilitates researchers to implement Human Activity Recognition (HAR) by employing deep learning techniques. Neural networks are becoming more popular in a range of scientific and educational research, and commercial applications. Neural network models are currently used in a diverse range of scientific, academic, and commercial applications to solve image processing problems. Deep learning-based frameworks, including 3D Convolutional Neural Networks, are capable of extracting spatiotemporal data from footage and detecting human activities. In this work, we propose a 3DCNN model to detect human activity from video sequences. The key research contribution is the pre-processing technique like key frame selection, background segmentation, modeling, and training of the efficient 3D Convolution Neural Network for classifying human activities. The model is tested on a wide variety of benchmark datasets like KTH, Weizmann, and UT-I. The efficacy of the suggested 3D CNN model in handling challenging or complex datasets was also evaluated. The proposed technique performs better than the reference methods in terms of recognition accuracy and training speed.
