A SlowFast behavior recognition algorithm incorporating motion saliency

Abstract

This paper first analyzes three major problems that can be encountered in video behavior recognition tasks: sampled blocks cannot be focused on motion regions, global motion affects recognition results, and the network's spatio-temporal modeling capability is weak. To address these three problems, we propose the SlowFast behavior recognition algorithm (MASlowFast) that incorporates motion saliency as an application scenario for mine personnel safety behavior recognition. The specific solutions are the sampling method based on motion saliency, the extraction of motion boundary features, and the spatio-temporal segmentation strategy of fast and slow channels. Finally, we validated the effectiveness and accuracy of the algorithm in this paper by ablation experiments on UCF101 dataset and HMDB51 dataset.