Research On Comprehensive Object Detection In Intelligent Monitoring Network

With the development of deep learning,object detection technology based on video images has become one of the main applications of monitoring equipment.But today there are still many challenges for efficient object detection in surveillance networks.First of all,the current monitoring network is mainly responsible for processing tasks by the central server.In a large-scale and high-resolution monitoring network,the excessive computing task volume will exceed the tolerance of the central server.At this time,by controlling the range and time range of the search,the task calculation amount can be effectively reduced to improve the processing efficiency.However,in a complex scenario,the amount of task calculation after narrowing the range may still be large,which requires fully utilizing the cameras in the monitoring network,so that the camera in the monitoring network has the capability of independently performing object detection.Although there are many mature object detection models at present,most of them are difficult to operate under the premise of ensuring accuracy and detection rate on development boards with limited resources.Therefore,the weight reduction of the model becomes the core of object detection at the cameras.In addition,when the cameras in the monitoring network independently perform object detection,due to the lack of communication and cooperation with others,an effective management mechanism on the serve is required for the scheduling management of each camera and the organic integration of the detection results.In view of the above situations and problems,combined with theactual scene and requirements of the monitoring network,this paper proposes a object detection technology based on comprehensive features,realizes independent object detection of the camera in the monitoring network.The cameras can process computing tasks together with the server.Finally the detection result of cameras will be organically fused on the server to obtain the location range of the object.This paper firstly gives the definition and representation of the comprehensive feature object detection technology.Comprehensive features include time,location,visual characteristics of the object,and the location characteristics of the cameras in the network.Then,according to the technical framework,the metadata management module,the task scheduling module,the task management module,the task execution module,the object detection module and the result fusion module of the server,the task execution module and the object detection module of the camera are introduced respectively.And the core technology in each module is elaborated.According to the input feature information,this paper uses the designed task scheduling algorithm to screen out the camera that needs to perform the object detection task,which reduces the waste of monitoring network resources while ensuring the accuracy.By cutting the SSD300x300 model,it achieves a lightweight effect,allowing it to run quickly on the camera development version,with a good detection speed.And through migration learning,each camera-side model has higher precision for the current scene.The task management module on the server side is responsible for maintaining the heartbeat communication with each camera,and making adjustments according to the working state of each camera and the detection result in time,and can be effectively processed when an abnormal situation occurs at the cameras.When the detection process is finished,the fusion result is designed by the fusion algorithm,and the two results of the granularity are used to fuse the detection results.The error correction algorithm is designed to reduce the error of the fusion result.The result determines the range of locations ofthe detectionobject.Finally,the CFOD system is realized with the comprehensive feature object detection technology as the core,and an analog monitoring network is built to verify and analyze the core technologies.It mainly includes the scheduling strategy accuracy and resource saving rate,the convergence speed,detection speed,accuracy and offloading task of the model,the feasibility and accuracy of the result fusion,and the abnormal situation.The experimental results show that through the comprehensive feature object detection technology,the camera task scheduling stage can save the resources of the monitoring network under the premise of ensuring the accuracy rate is above 93%;the detection speed of SSD300x300 model after the weight reduction on TK1 nad TX2 can reach to1.25 FPS and3.1FPS;the average accuracy is 84.8%;the average accuracy of the final fusion object range is 82.2%.