Abstract:

The taillight detection of the front vehicle is a research hotspot of environmental awareness in automatic driving. In order to detect the taillights of the vehicle in real time in a complex urban environment, Region-based Fully Convolutional Networks（R-FCN） is proposed into the taillight detection. A vehicle taillight detection algorithm based on multi-scale cascade R-FCN is proposed. Through the cross-layer connection in the network, the underlying features and high-level semantics of the taillights are merged, and the batch normalization layer is added to accelerate the convergence speed of the network, and an improved R-FCN sub-network is obtained. Then with a series of input data with different intersection-over-union, the trained R-FCN sub-networks are cascaded to obtain the final detection network. At the same time, the improved non-maximum suppression is utilized in inference stage to obtain the most accurate detection results. The test results show that the method achieves an overall mean average precision of 94.04% on the CVPR dataset, and the average detection time of a single image takes 31 ms, which has better performance in detection speed and accuracy.

Key words: vehicle taillight detection, Region-based Fully Convolutional Networks（R-FCN）, cascaded network, multi-scale feature fusion, batch normalization, non-maximum suppression

摘要：

前方车辆尾灯检测是自动驾驶中环境感知的研究热点，为在复杂城市环境下实时检测车辆尾灯，将基于区域的全卷积网络（Region-based Fully Convolutional Networks，R-FCN）引入尾灯检测，提出了一种基于多尺度级联R-FCN的车辆尾灯检测算法。通过网络中的跨层连接融合尾灯的底层特征和高层语义，并加入批次归一化层加快网络的收敛速度，得到改进的R-FCN子网络，将一系列在不同交并比输入数据上训练的R-FCN子网络级联得到最终的检测网络。同时预测阶段采用改进的非极大值抑制获得最精准的检测结果。检测结果表明，该方法在CVPR数据集上获得总体94.04%的平均精度，单张图片平均检测耗时31 ms，在检测速度和精度上均有较好的性能。

关键词: 车辆尾灯检测, 基于区域的全卷积网络（R-FCN）, 级联网络, 多尺度特征融合, 批次归一化, 非极大值抑制