Abstract: In view of the existing traffic situation of hybrid driving of autonomous vehicle (AV) and human-driven vehicle (HV) for a long time, a vehicle driving intention recognition model based on the taillight status in mixed traffic environment was proposed by combining deep learning and HSV color characteristic extraction method. The proposed model was limited by the region of interest (RoI) area detected by the Mask R-CNN vehicle detection model, and HV in this area was taken as the research object. According to the correlation of taillight positions, vertical and horizontal constraints were added in the rear area of HV to locate the taillights which transmitted the signal of light status. A taillight status recognition algorithm based on time sequence was proposed on the basis of the specified combination of light status and the flashing frequency of turn signals. The CSRT (channel and spatial relatability tracking) was used to track HV taillights and count the timing status of taillights. With the dynamical lights status as input, a driving intention recognition model based on taillight status was constructed. The Cityspaces dataset and traffic flow video data based on real road conditions were used to train, validate and test the proposed model. The research results show that the accuracy and recall rate for video stream vehicle taillight detection of the driving intention recognition model based on the taillight status are 96.0% and 98.2% respectively, the average accuracy rate of the driving intention recognition reaches 95.9%, and the recognition time of single frame is 20 ms. The proposed method provides an effective theoretical basis for AV decision-making and planning in high-speed mixed traffic environments.

Key words: transportation engineering; driving intention recognition; light status recognition; autonomous driving; Mask R-CNN; interactive behavior

摘要： 针对长期存在自动驾驶车辆(autonomous vehicle, AV)与人工驾驶车辆(human-driven vehicle, HV)混行的交通现状，结合深度学习和HSV颜色特征提取方法，提出了一种在混行交通环境下基于尾灯灯语的车辆驾驶意图识别模型。以Mask R-CNN(mask region proposal convolutional neural network)车辆模型检测出的感兴趣区域RoI(region of interest)为限制，将区域内的HV作为研究对象，根据尾灯位置相关性，在HV车尾区域添加纵横向约束来定位传递灯语信号的左右尾灯；在规定的灯语组合及转向灯闪烁频率基础上，提出了一种基于时间序列的灯语识别算法，运用多目标判别相关性滤波CSRT(channel and spatial relatability tracking)跟踪HV尾灯并统计尾灯时序状态；以动态灯语作为输入，构建基于尾灯灯语的驾驶意图识别模型；使用基于真实路况信息的Cityspaces数据集和交通流视频数据对模型进行训练、验证和测试。研究结果表明：基于尾灯灯语的驾驶意图识别模型对视频流车辆尾灯检测准确率和召回率分别为96.0%、 98.2%，对驾驶意图识别的平均准确率达到了95.9%，单帧识别耗时为20 ms，为高速混行环境下的AV决策规划提供了有效的理论依据。

关键词: 交通工程；驾驶意图识别；灯语识别；自动驾驶；Mask R-CNN；交互行为