Spatio-Temporal Attention Fusion SlowFast for Interrogation Violation Recognition

The use of video surveillance to monitor interrogation behavior can effectively maintain judicial civility in the context law enforcement cases. However, analyzing and reviewing videos be a time-consuming resource-intensive process, particularly manual identification violations. This work is dedicated development an intelligent recognition system for violations by using spatio-temporal attention fusion SlowFast Network. To address issue feature information underutilization slow path traditional SlowFast, slow-to-fast incorporated into original enhance learning. model fuses spatial temporal channels, replacing convolution module with this new approach. proposed was evaluated publicly available UCF101 action dataset, resulting 1.52% improvement Top-1 accuracy compared SlowFast. Based on two custom misconduct datasets, evaluated, achieving rate 99.16% misconduct. demonstrates its effectiveness identifying behaviors inside rooms. Compared some advanced models, strong competitiveness during interrogations.