Research On And Implementation Of Object Detection Algorithms Based On Regional Security

In recent years,with the construction of intelligent cities and safe transportation in China,the industry involving security technology has become an important part of economic construction in the new era.Traditional security technology is difficult to meet the needs of society in terms of safety and convenience,and the visual security system based on deep learning is gradually adopted by people.Although the visual security technology improves the detection efficiency and accuracy,in some scenarios,it is still difficult to detect the targets accurately in real time.Although in most case,missed detection,wrong detection and untimely detection may have little impact on the security area,once there exist important objectives in missed detection,wrong detection and untimely detection,it will pose a huge threat to the security area.Therefore,it is necessary to monitor the area in real time and accurately,so as to reduce the potential security hazards in the monitoring area.This paper mainly studies the object detection algorithm in the field of visual security technology.Using the method of deep learning and taking a large number of open image and video data sets as the research object,the target detection method on surveillance video is studied and optimized.Combined with engineering application,the real-time target detection of security surveillance video is realized,which has certain innovation.The main work and research results are as follows: 1.In view of the problem that overfitting is prone to occur in VGG network structure and SSD detector is difficult to detect weak targets.In this paper,an improved RSSD target detection model based on SSD is proposed,which reduces the number of parameters in the network structure and improves the ability of the model to extract image features.The experimental results show that the proposed RSSD target detection model outperforms the SSD model and improves the detection accuracy of the target.Moreover,the RSSD model is more sensitive to the detection of small targets in images.2.In view of the phenomenon that objects often occlude each other in the visual security scene,resulting in false detection and missed detection.In this paper,a novel loss function based on variant IOU is proposed to replace the regularized loss function in the existing model.A new loss function is used to better close the distance between the prediction frame and the real frame of the same detection object and reduce the false detection rate.The comparison experiment proves that the proposed variant IOU loss function is suitable for the existing object detection model,which improves the robustness of the model to the occlusion of the object and the detection accuracy.3.For the application of visual security technology in engineering,the requirement of realtime monitoring for hardware equipment is relatively high,which leads to the lack of popularity of applications.In this paper,an RSSD-IOU object detection algorithm is implemented based on Jeston TX2 hardware.The compressed RSSD model and variant IOU loss function are integrated and mounted on Jeston TX2 hardware.Experiments show that the detection speed of the model on Jeston TX2 can meet the real-time performance and has stable performance.