HOI-Based Method for Detecting Cell Phone Behavior in Industrial Safety Production Environments

To address the practical problems of algorithm inference timeliness and the complex environment in industrial settings, The presented dissertation introduces a comprehensive human-object interaction recognition technique, leveraging the attention-based architecture of the Transformer model. By sampling key features at different scales from the industrial complex environment and using Transformer's attention mechanism to aggregate global information, the algorithm achieves accurate detection of targets in complex environments while eliminating post-processing processes such as non-maximal value suppression to achieve real-time detection. Moreover, a deformable attention module is used to optimize small target detection and the multi-headed attention mechanism of Transformer, which adaptively adjusts the position and shape of the feature extraction window, reducing the computational effort of the attention computation process and accelerating the model training time. The investigation outcomes reveal that the introduced HOI-centric technique for identifying mobile phone infractions within industrial safety contexts exhibits superior accuracy in detection and notable enhancements in computational efficiency when benchmarked against other sophisticated algorithms in the expanded mobile phone category within the HICO-DET dataset.