Abstract: For intelligent vehicles, if the sensing device might accurately and quickly detect the concave and convex obstacles on the roads ahead of the vehicles, the important preview information might be provided for the control of the chassis system such as the suspension of the vehicles, and finally realized the improvement of the comprehensive performance of the vehicles. Therefore, based on improved YOLOv7-tiny algorithm a recognition method was proposed for typical positive and negative obstacles such as bumps(speed bumps) and pits on the road surfaces. Firstly, the SimAM module was introduced in the three feature extraction layers of the original YOLOv7-tiny algorithm to enhance the networks ability to perceive the feature map; secondly, a smoother Mish activation function was used in the Neck part to add more nonlinear expressions; again, replacing the nearest proximal upsamping operator with the up-sampling operator to enable the network to aggregate contextual information more efficiently; and lastly, the WIoU was used as the localization loss function to improve the convergence speed as well as the robustness of the network. The offline simulation experimental results show that compared with the original model, the improved model improves the average accuracy by 2.5% for almost the same number of parameters with an intersection ratio of 0.5 between the predicted and real frames. The improved model is deployed to a real vehicle, and the real-vehicle experiments verify that the model may effectively detect the obstacles appearing on the road in front of the vehicles, indicating that the proposed algorithmic model may accurately provide the pre-precedent information for obstacle detections.

Key words: road preview, uneven obstacle, improvement and optimization, recognition method

摘要： 对于智能车辆而言，如果感知设备可以准确快速地检测到车辆行驶前方道路上的凹凸障碍物，那就可为车辆悬架等底盘系统的控制提供重要的预瞄信息，从而实现车辆综合性能的提高和改善。针对路面上常见的凸块（减速带）、凹坑等典型的正负障碍物，提出了一种基于改进YOLOv7-tiny算法的识别方法。首先，在原YOLOv7-tiny算法的三个特征提取层引入SimAM模块，增强网络对特征图的感知能力；其次，在Neck部分采用更为平滑的Mish激活函数，增加更多的非线性表达；再次，使用CARAFE上采样算子替换最近邻上采样算子，使网络更有效地聚合上下文信息；最后，将WIoU作为定位损失函数，提高网络收敛速度以及鲁棒性。离线仿真实验结果表明：与原模型相比，改进后的模型在几乎相同参数量下，预测框与真实框交并比为0.5时的平均准确度提高了约2.5%。将改进后的模型部署到实车上，实车实验验证了模型能够有效检测出车辆前方路面出现的障碍物，说明所提出的算法模型能够准确提供障碍物检测的预前信息。

关键词: 路面预瞄, 凹凸障碍物, 改进与优化, 识别方法