Research On Moving Target Detection And Unsafe Behavior Recognition In Power Grid Construction Site

With the rapid development of China’s power grid,the corresponding power infrastructure construction is gradually increasing,and the construction safety under the power grid environment is increasingly concerned.The construction site of power grid is complicated,so it is necessary to timely detect and find the potential safety hazards.In this paper,the possible causes of accidents in power grid construction site are summarized into two parts: the wearing condition of personnel’s safety helmet and the operation condition of engineering vehicles on site.Based on the wide application of monitoring equipment and target detection algorithm,the moving targets and unsafe behaviors in power grid construction site are studied.The main work is as follows:(1)Whether the mandibular strap is fastened is taken as the basis of whether the safety helmet is properly worn.Thus,the image sample database including "correctly wearing the safety helmet","incorrectly wearing the safety helmet" and "not wearing the safety helmet" is constructed.The YOLOv3 algorithm is used for example experiments and compared with the Faster R-CNN,SSD and YOLOv4 algorithms.The results show that the mean average precision(m AP)and frames per second(FPS)of YOLOv3 are 93.52% and 17.75 respectively.YOLOv3 has better detection performance by combining detection accuracy and speed.(2)By using Grab Cut image segmentation and Poisson fusion algorithm,the construction vehicle image database including "crane","truck","bulldozer" and "excavator" is constructed;Replacing YOLOv3 backbone feature extraction network with lightweight network Mobilenetv3 to build Mobilenetv3-YOLOv3 model;By adding attention mechanism SENet on the basis of YOLOv3 original network to build SENet-YOLOv3 model.According to comparative experiments,the m AP value of two model are 70.97% and 95.31% respectively,and FPS value are 33.94 and 23.45 respectively,indicating that the latter has better detection effect.Comparing with Faster R-CNN,SSD and YOLOv4 algorithms,it is further proved that SENet-YOLOv3 is more suitable for engineering vehicle detection.(3)Using the single image and the actual video of the construction verify the two detection model,the results showed that the safety helmet wearing detection model can effectively detect whether personnel are wearing helmets in substation equipment maintenance operations,line inspection and repair operations,high-altitude line operations,and it also has a good detection effect in the multi-person operation situation in the actual video;The construction vehicle detection model can correctly identify construction vehicles under construction,and can continuously and accurately detect distant vehicle targets in actual videos.