基于DCN-Mobile-YOLO模型的多车道车辆计数

Multi-lane vehicle counting based on DCN-Mobile-YOLO model

单一目标检测方法无法实现目标计数的准确统计,且模型的检测精度和速度难以同步提升.以YOLO v4目标检测框架为基础,提出一种移动端的目标追踪和多车道车辆计数模型DCN-Mobile-YOLO.使用可变形卷积网络(deformable convolutional networks,DCNs)v2卷积核和移动端卷积网络MobileNet v3框架分别代替YOLO v4的常规卷积核和主干网络,结合DeepSORT算法实现对多目标的跟踪和计数,建立自适应车道检测规则并实现车道内车辆的精确计数.在VOC2007+2012数据集和GoPro采集数据上验证DCN-Mobile-YOLO模型的有效性.结果表明,DCN-Mobile-YOLO模型的平均精度均值相比主干网络为MobileNet v3和CSPDarkNet的YOLO v4算法分别提升了13.19%和6.63%,目标检测平均帧率为12帧/s.DCN-Mobile-YOLO模型不仅提高了目标检测模型的检测精度,且达到了移动端实时检测的速度.

The single object detection method cannot achieve the accurate statistics of object counts and the accuracy and speed of model detection cannot be improved simultaneously.To solve these problems,we propose a mobile-side object tracking and multi-lane vehicle counting method-DCN-Mobile-YOLO based on the YOLO v4 object detection framework.This method uses the DCNs v2 convolution kernel and MobileNet v3 framework to replace the conventional convolution kernel and backbone network of YOLO v4,respectively,and uses the DeepSORT algorithm to achieve multi-object tracking and counting.At the same time,an adaptive lane detection rule is proposed to achieve the accuracy of vehicles in the lane count.Finally,the effectiveness of DCN-Mobile-YOLO method is verified on the VOC2007+2012 data set and GoPro collected data set.The experiment results show that the mAP of DCN-Mobile-YOLO method is improved by 13.19%and 6.63%respectively compared with YOLO v4 algorithm of MobileNet v3 and CSPDarkNet as the backbone network,and the average frame rate of object detection is 12 frame/s.Our DCN-Mobile-YOLO method not only improves the detection accuracy of the object detection model but also achieves the speed of real-time detection on the mobile terminal.