基于改进SSD的无人驾驶夜间目标检测

Unmanned night object detection based on improved SSD

为提高SSD(single shot multibox detector)网络在夜间无人驾驶的目标检测性能,对SSD网络进行了改进:特征提取网络使用稠密连接卷积网络(Densenet)得到表征能力更强的特征图;卷积过程中进行特征图重利用,从而增加中小目标的信息;加入反卷积网络丰富特征图语义信息。试验的训练方法采用无预训练模型的训练方式,这种训练方式能够更好地拟合数据集的特性。试验结果表明,改进的SSD网络的目标检测准确率高于原SSD网络,并且优于其他主流网络。

Single shot multibox detector(SSD)network was modified to improve the performance of target detection of unmanned driving at night.Feature extraction network used densely connected revolutionary networks(Densenet)to obtain more powerful feature maps.In the process of convolution,the feature map was reused to increase the information of medium and small targets.Deconvolution network was added to SSD network to enrich semantic information of feature map.The experimental training method adopted the training method without pre-training model,which could better fit the characteristics of data set.The experimental results show that the target detection accuracy of the improved SSD network is higher than that of the original SSD network,and better than other mainstream networks.