Spatio-Temporal Attention Fusion SlowFast for Interrogation Violation Recognition

Abstract

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