Low-High Feature Based Local-Global Attention for Traffic Police Action Identification

Conventional Convolutional Neural Networks (CNN) provide the inherent capability to effectively capturing the local area of the data. In order to understand a human action, it is necessary to consider both human and the overall context of given scene. Hence, we incorporate an attention mechanism into our model to enhance the retrieval of global semantic context during action recognition. This study introduces novel network designs that employ the Local-Global attention mechanism to extract comprehensive characteristics from video data. The initial step involves the utilization of a 3D Convolutional Neural Network (3DCNN) to extract both structural and semantic information from RGB inputs. The second approach involves enhancing the understanding of visual information by utilizing an attention mechanism to refine spatial configurations. By placing emphasis on relevant components, this approach is expected to facilitate the distinction of actions. Specifically, the Spatial-Attention (SA) generation layers priorities the visual features of the subject, thereby highlighting the pertinent scene background information surrounding the objects. The evaluation of our proposal is conducted in a comprehensive manner, utilizing the datasets provided by the Traffic Police. In the experiments, the results reveal that the proposed DNN achieves the average accuracies of 97.6
^%
in Police Office the best of our knowledge.