Research On Smoking Behavior Detection Method For Outdoor Constructors In Mid-distance

In recent years,the usage rate of video surveillance in all professions and trades has been increasing rapidly.People's functional requirements for monitoring equipment are not just simply recording.Using computer vision and image technology to help analyze the target's behavior in surveillance so as to realize intelligent surveillance,is the future development direction of video surveillance industry.There are a great number of intelligent surveillance equipments on market currently.Most of them are applicable for indoor scenes;few products that can apply to outdoor scenes,can only implement monitoring and behavior detection in a very short distance.Considering the background,this dissertation started from the perspective of abnormal behavior detection of construction workers in outdoor construction scenarios.A detection algorithm run on embedded devices with limited computing power,which can detect the typical abnormal bevavior smking in the range of 5~10 meters by using surveillance video clips,is the result of this dissertation.The detection algorithm solves two core problems.One is the detection and segmentation of human body in surveillance video clips,and the other is the detection and recognition of smoking behavior in outdoor construction scenarios.When implement the detection and segmentation algorithm of the human body in the surveillance video clips,this dissertation designs and proposes a human detection and segmentation algorithm called viewfinder mode.The core thought of the algorithm is to try to find a rectangular segmentation frame corresponding to the target person,and then use this segmentation frame to intercept the person's image in the surveillance clips.Because the same rectangular segmentation frame is used for the interception of each frame of surveillance clips,the person in the segmentation result looks just like he is in a viewfinder,which is why this human body detection segmentation algorithm was named the viewfinder mode.Using this solution,while ensuring the accuracy of human segmentation,reduces the occupation of embedded device computing resources,thus improving the efficiency of the detection algorithm.As for the detection algorithm for smoking behavior,this dissertation designs a twostep detection plan,which is consisting of primary screening and secondary screening.The core thought of the plan is to divide the process of the detection into two stages,and then detecting one's behavior from the perspective of target detection and behavior detection in the two stages.With the two-step plan,behaviors which are definitely not smoking can be excluded as much as possible in the primary screening stage,so that more computing resources of the embedded device are reserved for the detection of suspicious smoking behavior in the secondary screening stage.The two-step detection plan is the core of this dissertation,which makes it possible to reconcile the reduced performance requirements of embedded device and the accuracy of the detection algorithm.Compared with other algorithms,the detection algorithm in this dissertation is designed to run on those low-cost and easy-to access embedded devices represented by Raspberry Pi.Besides,the detection algorithm is applicable for outdoor construction scenarios.Therefore,this detection algorithm is more practical.The evaluation shows that the accuracy of the detection algorithm is 86.59%,which is expected.