Activity Recognition Using Temporal Features and Deep Bottleneck 3D-ResNeXt

Abstract

Due to the advancements in the field of artificial intelligence, object detection and classification are becoming easier. New deep learning and machine learning-based approaches are being used for video analysis, activity recognition and activity detection, too. But, training of deep learning-based neural networks for large video datasets is computationally expensive. Also, it lacks in terms of either speed or correctness and has drastic variations due to the background, scale, etc. Thus, while deep learning has given outstanding results for image classification and detection, activity detection and recognition can still use better approaches for faster and more accurate results. In this paper, we built a model based on ResNet classifier and ResNeXt classifier, which gives improved performance in action recognition. We start with introductory topics and network architecture. Then, we discuss the experiments that we have tried, and also, we show the process of how we get to deploy this method in edge devices like Nvidia Jetson Nano and Intel UP Squared board.