基于双主干网络的雾天交通目标检测方法研究

车辆和行人安全监测是城市交通监测的一项重要任务。 针对雾霾等复杂恶劣天气条件下,监测采集的图像视觉效果差、噪声高、目标检测困难等问题,提出了一种双主干网络(MobileNets VGG-DCBM Network, MVNet)用于雾天交通目标检测,结构受 PCCN 和 CBNet 网络结构的启发,由改进的深度可分离卷积神经网络 MobileNets 和基于 VGGNet 构建的 VGG-DCBM 网络组成;采用并行方式构建双主干目标检测网络结构,以改进的 MobileNets 为主主干网络,VGG-DCBM 为辅助主干网络,共同提取特征信息,实现不同网络间特征层信息的融合;MVNet 网络结构采用并行方式获取两个不同网络提取的不同特征层信息,通过采用通道拼接的方法实现不同网络特征信息之间的融合,以获得更丰富的细节特征;在 RTTS 和 HazePerson 数据集上,平均精度均值(mean Average Precision, mAP)分别达到 71. 50%和 89. 84%;实验结果表明:在雾霾等复杂恶劣天气条件下具有较强的鲁棒性且能够准确的检测到车辆和行人,在目标检测性能上优于对比方法。

Research on Traffic Object Detection Method in Fog Based on Dual Backbone Network

Vehicle and pedestrian safety monitoring is an important task of urban traffic monitoring. Aiming at the problems of poor visual effect high noise and difficult target detection of the images collected under complex and bad weather conditions such as fog and haze a dual backbone network MobileNets VGG-DCBM Network MVNet was proposed for traffic object detection in fog. Inspired by the network structure of PCCN and CBNet this structure was composed of improved depthwise separable convolutional neural network MobileNets and VGG-DCBM network based on VGGNet. The dual backbone object detection network structure was constructed by using the parallel method. The improved MobileNets was the main backbone network and VGG-DCBM was the assistant backbone network to extract the feature information and realize the fusion of feature information between different networks. MVNet network structure adopted the parallel method to obtain the information of different feature layers extracted by the two networks and realized the fusion of different network feature information by using the method of channel splicing so as to obtain richer detailed features. On RTTS and HazePerson datasets the mean average precisions mAP reached 71. 50% and 89. 84% respectively. The experimental results show that this method has strong robustness in complex bad weather conditions such as fog and haze and can accurately detect vehicles and pedestrians. It is better than the comparison method in object detection performance.