Temporal action proposal generation Based on Boundary Probability

Temporal action proposal generation is widely used in video processing, action recognition and other fields. Aiming at the problems of the existing methods of temporal action proposal generation that fail to effectively use the Temporal information of video frames and the blurring of action boundaries, this paper proposes a method of time-series action proposal based on boundary probability, and the existing method based on slider and action score temporal action proposal generation method is different. In this paper, a time series boundary probability generation network is generated to generate the boundary sequence in order to optimize the time series boundary fuzzy problem; at the same time, the proposal evaluation network is constructed to optimize the time series proposal segment; in addition, the improved Maximum value suppression algorithm--DTS-NMS to further improve the temporal action proposal generation results. The experimental results show that the constructed model has achieved competitive effects on the two public datasets of ActivityNet-1.3 and THUMO14 (represented as large datasets and small and medium datasets, respectively).