Multi-Scene Safety Helmet Detection with Multi-Scale Spatial Attention Feature

The safety helmet detection system based on video surveillance has appeared in many smart construction sites. However, existing safety helmet detection algorithms have difficulty in detecting overlapping and small objects owing to the influence of complex environmental, and the features of safety helmet contain noise unrelated to the detection object which resulting poor detection performance. To address this problem, in this paper, a layer feature weighted module (LFWM) is added after different scale feature maps and getting the score matrix of the same size as the feature map. Finally, point-wise multiply is applied between score matrix and feature map for filtering the irrelevant noise. This method can highlight the local features of safety helmets in different feature maps and suppress the noise features that are not related to the detection object. Experiments show the proposed method can improve the safety helmet detection performance in different scenarios, and the (mean Average Precision) mAP improved by 4.19% compared with the original RetinaNet (ResNet50).