Wearing Detection by Multi-Scale Information Fusion Network for Live-Line Working Scenarios
Junyi Zhou, Yanduo Zhang, Tao Lü, Wenhua Fang · 2022
In live-line working scenarios, properly wearing the insulating protective personal equipment (PPE) is an important means of ensuring the safety of the operator. However, there are many challenges in applying universal object detection for wearing detection in live-line working scenarios. Different cameras are located at different distances from the operator. During the operation, the insulating bucket where the operator stands will keep moving. These can lead to a large range of scale changes of objects in the image, causing a decrease in detection accuracy. In order to solve the above problems, we design a PPE dataset about live-line working and propose a multi-scale information fusion network based on Yolov5. To be specific, in order to suppress the inconsistency between different scales, we employ the adaptive spatial feature fusion module. Then we integrate coordinate attention to make the network pay more attention to the location information, alleviating the problem of small target miss detection. Moreover, we also adopt Varifocal loss to mitigate the category imbalance problem to improve accuracy. The experiments prove that our method outperforms most object detection methods. And the results show that our multiscale information fusion network achieves 90.3% AP50 and 61.4% AP50:95 on our dataset. Compared to Yolov5, our method improves 2.2% AP50 and 2.1% AP50:95. For the shawl category, AP50 improves by 7.4%.