基于改进YOLOv3网络的卡尔曼社交距离检测与追踪

为了预防新冠肺炎的传播，在佩戴口罩的同时，保持一定的社交安全距离是必要的。为解决现有的目标检测算法在社交距离检测中无法同时满足检测的实时性、准确性以及在复杂场景中存在遮挡、小尺度目标等问题，提出基于YOLOv3的改进算法DPPY(Dilated Pyramid-Pooling with YOLOv3)。首先使用空洞卷积参与到YOLOv3的核心图像处理结构中，然后引入密集型连接网络进一步融合不同层之间的连接，并且在这基础上还模仿了空间金字塔结构处理输入数据的尺寸问题，最后将这些处理结果一起与待追踪物体与彼此间的前后位置进行更好的关联并选用卡尔曼滤波器这个工具来更好地处理。若行人彼此间靠的过于紧密，则标红发出警报，以便更好地提醒相关人员注意。结果表明：与传统的YOLOv3算法相比，DPPY算法检测速度更快，检测精度更高。检测速度达到了34帧/s，平均准确率(Average Precision, AP)提高了9.1 %，并且在大、中、小目标检测中平均准确率均值(mean Average Precision, mAP)分别提高了7.8 %、8.2 %、8.9 %。

Social Distance Detection and Tracking Algorithm Based on Improved YOLOv3 and Kalman Filter

In order to prevent COVID-19 from spreading, it is necessary to maintain a certain social security distance while wearing a mask. In order to solve the problems that the existing object detection algorithms can not meet the real-time and accuracy of social distance detection at the same time, and there are occlusion and small-scale objects in complex scenes, an improved algorithm DPPY (Dilated Pyramid-Pooling with YOLOv3) based on YOLOv3 is proposed. Firstly, use the dilated convolution to participate in the core image processing structure of YOLOv3, and then introduce a dense connection network to further merge the connections between different layers, and on this basis, it also imitates the spatial pyramid structure to deal with the size of the input data, and finally these processing results are better correlated with the objects to be tracked and the front and back positions of each other, and the Kalman filter tool is selected for better processing. If the pedestrians are too close to each other, the warning will be issued in red to better remind the relevant personnel to pay attention. The results show that DPPY algorithm has faster detection speed and higher detection accuracy than traditional YOLOv3 algorithm. The detection speed reaches 34 frames per second, the average precision (Average Precision, AP) is increased by 9.1 %, and the mean average precision (mean Average Precision, mAP) is increased by 7.8 %, 8.2 %, and 8.9 % in large, medium and tiny target detection, respectively.