【作者】 孙迎春；

【导师】 潘树国；

【作者基本信息】 东南大学 ， 精密仪器及机械， 2022， 硕士

【摘要】 道路环境检测技术能够为智能车辆实现高效决策和精准控制提供丰富动态的周围环境信息,是保障车辆安全行驶的基础和关键。而实际驾驶过程中交通场景复杂、道路结构多变、目标类型众多,使得智能车辆道路环境检测技术仍然面临诸多挑战,准确性和实时性有待进一步提高。对此,本文基于计算机视觉技术和深度学习算法开展道路环境检测技术研究,旨在利用深度卷积神经网络模型对复杂交通场景下包括道路目标、可行驶区域在内的道路环境信息进行检测,设计并提出优化算法。主要研究内容如下:(1)提出了基于AFEM-YOLO的道路目标检测算法。针对YOLOX提取的行人、车辆等道路目标特征包含干扰和冗余信息的问题,设计一种注意力特征增强结构,利用注意力机制为网络参数更新过程中的通道赋权提供指导和选择性地聚合区域信息,以实现高效有针对性的融合低层空间信息和高层语义信息。在此基础上,引入语义增强分支以进一步提升网络的整体特征层次,提高模型在复杂交通场景下对道路目标的检测性能。在公开数据集PASCAL VOC 2007和BDD100K上的实验结果表明,注意力特征增强模块实现了对多类道路目标特征的有针对性筛选与自适应细化,AFEM-YOLO分别取得了84.1%m AP和59.0%m AP,较原YOLOX算法提高0.6%和1.8%。(2)提出了基于GH-Road Seg的可行驶区域检测算法。针对Road Seg算法因引入深度信息提高可行驶区域检测性能而导致参数数量增加的问题,设计全局注意力模块以加强网络对全局特征的提取和关注,充分利用全局特征提供的上下文信息为像素级分类服务。此外,构造分层特征融合金字塔结构,利用其丰富深层特征的图像结构信息以获取深层高分辨率特征映射,降低编码网络与解码网络的连接通道与支路,达到优化可行驶区域分割结果的同时有效降低模型参数数量的目的。在公开数据集KITTI和Cityscapes上的实验结果表明,GH-Road Seg算法较原Road Seg算法的可行驶区域检测性能分别提升了0.6%和0.5%,参数量降低了17%。(3)提出了基于Multi-YOLO的多任务联合检测算法。针对分别运行道路目标检测与可行驶区域检测两个独立网络模型的资源开销大、实时性低,模型间输入输出难以同步的问题,在前述两章算法的基础上设计基于编码-解码结构的多任务网络模型。MultiYOLO利用编码器提取共享输入特征,在捕捉任务间关联性的同时减少模型参数,提高运行效率,并通过不同的解码分支学习单个任务的差异性。在公开数据集BDD100K上的实验结果表明,Multi-YOLO在道路目标检测任务和可行驶检测任务分别取得了77.1%m AP和91.9%m IOU,同时与单独执行每个任务相比,该算法将模型推理速度提高33.5%。更多还原

【Abstract】 Road environment detection technology can provide abundant and dynamic surrounding condition information for intelligent vehicles to achieve efficient decision-making and precise control,which is the basis and key to ensure the safe driving of automobiles.However,in the actual driving process,the complex traffic scenes,changeable road structures and various types of objects make the road environment detection technology of intelligent vehicles still face plenty of challenges,and the accuracy and real-time need to be further improved.Therefore,this paper studies the road environment detection technology based on computer vision technology and deep learning algorithm,aiming at detecting road environment information including road objects and drivable area in complex traffic scenes by using deep convolution neural network model,and designing and proposing corresponding optimization algorithms.The main research contents are as follows:(1)The road object detection method based on AFEM-YOLO is proposed.Aiming at the problem that the pedestrian,vehicle and other road object features extracted by YOLOX contain interferential and redundant information,an attention-based feature enhancement module is designed.It utilizes the attention mechanism to provide guidance for channel weighting in the process of network parameter updating and selectively aggregate regional information,so as to achieve efficient and discriminating fusion of low-level spatial information and high-level semantic information.On this basis,semantic enhancement branch is introduced to further promote the overall feature level of the network and enhance the detection performance of the model for road objects in complex traffic scenes.The experimental results on the public datasets PASCAL VOC 2007 and BDD100 K show that AFEM realizes the targeted screening and adaptive refinement of diverse features,and AFEM-YOLO achieves 84.1%m AP and59.0%m AP respectively,which is 0.6% and 1.8% higher than the original YOLOX algorithm.(2)The driveable area detection algorithm based on GH-Road Seg is presented.Considering the problem that the number of parameters of Road Seg algorithm increases due to the introduction of depth information to improve the detection performance of driveable area,the global attention module is designed to strengthen the extraction and concern to the global features,and make full use of the context information to serve the pixel-level classification.In addition,a hierarchical feature fusion pyramid module is constructed,which can enrich the image structure information of deep features to obtain deep high-resolution feature mapping,and reduce the connecting channels and branches between coding and decoding network,in order to optimize the segmentation results of drivable area and effectively decrease the number of model parameters.Experimental results on public datasets KITTI and Cityscapes indicate that GH-Road Seg boosts the detection performance of the driveable area by 0.6% and 0.5%respectively,and diminishes the number of parameters by 17% compared with the Road Seg.(3)The multi-task joint detection algorithm based on Multi-YOLO is constructed.To solve the problems of high resource cost,low real-time performance and difficult synchronization of input and output between two independent network models of road object detection and driveable area segmentation respectively,a multi-task network model based on encoderdecoder structure is designed on the basis of the algorithms in the above two chapters.MultiYOLO uses encoder to extract the shared input features,which can capture the correlation between tasks,reduce the model parameters and improve the inference efficiency,and learn the differences of individual tasks through different decoding branches.The experimental results on the public dataset BDD100 K manifest that Multi-YOLO has achieved 77.1% m AP and 91.9%m IOU in road object detection task and driveable area detection task respectively,and compared with performing each task separately,the multi-task network model can enhance the inference speed of the model by 33.5%.更多还原