Research On Detection Method Of Dangerous Behavior Of Personnel In Port Environment

With the rapid development of computer vision technology,the method of detecting dangerous behaviors in videos using computer vision technology has been widely applied.Although the YOLOv5 object detection algorithm has significant advantages compared to other object detection algorithms in the same period,there are still certain shortcomings in the YOLOv5 object detection algorithm in detecting dangerous behavior of port personnel.This article conducts in-depth research on the detection of three behaviors: wearing masks,making phone calls,and smoking based on the YOLOv5 object detection algorithm.The main work of this article is as follows:Firstly,in response to the complex port environment and the difficulty in extracting effective features of masks and mobile phones,a mask wearing and phone making detection algorithm based on regional maximum expansion convolution is proposed.We proposed the region maximum dilated convolution,which is a new type of convolution method that greatly reduces the impact of complex backgrounds on target objects and enhances the transmission of feature information between different network layers;Introducing Anchor Free to replace the Anchor Base in the original YOLOv5 solves the problem of pre-set initial boxes not being able to match the target object well.The initial box no longer needs to be pre-set,but rather generates more suitable anchor boxes for the target object through autonomous learning of the network model.The experimental results show that the mask wearing and phone call detection algorithm based on region maximum dilated convolution proposed in this paper has greatly improved the detection performance.Secondly,a smoking detection algorithm based on enhanced multi-scale feature fusion and attention mechanism is proposed to address the problem of difficulty in retaining key feature information when smoke belongs to small target objects.Although YOLOv5 has good detection performance in detecting medium to large target objects,there are still some shortcomings in the accuracy of detecting small target objects.Therefore,it is proposed to strengthen the multi-scale feature fusion structure,so that each layer of the network can obtain more rich feature information of small target objects.By introducing CA attention mechanism,the feature layer used for prediction in the network can adaptively focus on small target objects.The experimental results indicate that the smoking detection algorithm proposed in this article based on enhanced multi-scale feature fusion and attention mechanism has made significant progress in the detection of smoking.Finally,the two detection algorithms proposed above will be experimented on their respective datasets,and the experimental results will be analyzed in detail.A dangerous behavior detection system for port personnel has been constructed,which achieves real-time dangerous behavior detection and alarm for port personnel.