Unsafe Behavior Detection In Construction Area Based On Deep Learning

Accidents caused by unsafe behavior of construction personnel often occur in the construction area of enterprise's mining and production process.Once the accident happens,the impact on the normal production of construction personnel and enterprises is catastrophic.Therefore,in order to prevent the occurrence of safety accidents caused by unsafe behavior of workers in construction area,it is of great significance to study the real-time detection and recognition of unsafe behavior of workers in construction area.In this thesis,the detection and identification of the two unsafe behaviors of the staff wearing incorrect safety helmet and smoking at the construction site are studied in depth.Considering that deep learning algorithm has great advantages over traditional pattern recognition,this thesis uses the method of deep learning to do related research.And the following work has been done:1)The general DSOD Deep Learning Detector is improved to make it suitable for the detection of unsafe behavior in this thesis.Because the general purpose DSOD target detector usually detects multiple different categories.At the same time,these categories are different and mutually exclusive.However,there may be only one object to detect unsafe behavior in this thesis,but there may be multiple unsafe behaviors in the detected object at the same time.That is to say,the detected categories in this thesis are actually not mutually exclusive,and there is not only one category.Therefore,this thesis extends DSOD to multi-label detection.For this requirement,this thesis finally replace the softmax layer,using SIGMOD as the active function of the last layer.The original classification task is transformed into multiple binary classification tasks,so that the detector can adapt to the detection task in this thesis.2)In view of the non-mutually exclusive nature of unsafe behavior,this thesis designs a loss function suitable for the detection of unsafe behavior.In order to further improve the performance,Focal loss is introduced into the loss function.3)For the DSOD network structure,this thesis uses the SE-block for its further improvement.In order to achieve better detection results,SE-block in SENet is embedded in DSOD's Dense block,and the network structure is separated between classification and boundary box regression,which reduces the impact of the difference between the two tasks.4)Aiming at the problem that the detection speed of the model may be insufficient,this thesis presents a compression scheme based on channel separation for the detection model.Experiments on self-built unsafe behavior datasets verify the feasibility of the unsafe behavior detector designed in this thesis and the effectiveness of the model improvement method.