基于毫米波雷达和相机融合的无人碾压机施工障碍物快速精准感知方法

A rapid and accurate obstacle perception method for unmanned roller in construction based on millimeter-wave radar and camera fusion

对施工环境的快速精准感知是保证无人碾压机安全、稳定运行的基础。然而当前坝面碾压施工过程中无人碾压机仅依靠毫米波雷达对障碍物的距离进行感知,当距离小于给定阈值时,碾压机由于无法感知障碍物的类别,常引起障碍物的误识别,导致无法进行后续碾压作业。针对上述问题,本研究提出基于毫米波雷达和相机融合的无人驾驶碾压机施工障碍物快速精准感知方法。首先,该方法将快速区域卷积神经网络(Faster R-CNN)特征提取网络中的卷积运算核替换为不同扩张率的空洞卷积核(DC),实现对坝面障碍物类型的快速精准识别。然后,将毫米波雷达感知到的障碍物距离、速度信息与相机感知到的类别信息进行融合,实现对坝面环境的全面精准感知。工程案例表明,相较于现行的Faster R-CNN目标检测算法,本研究提出的DC-Faster R-CNN目标检测算法mAP(检测各类障碍物的平均精度值)提高了2.59%,每张图片的检测时间减少了0.015 s;同时,基于多元信息融合的感知策略实现了碾压施工过程中的精准避障,保证了坝面施工的安全和效率。

The rapid and accurate perception of the construction environment by unmanned roller is essential for ensuring the safe and stable operation of the unmanned roller.However,in the current process of dam roller compaction process,unmanned rollers rely solely on millimeter-wave radar to perceive the distance to obstacles.When the distance is less than a given threshold,the roller stops and waits.This approach fail to identify the category of obstacles,frequently resulting in the misidentification of obstacles,resulting the roller waiting for the subsequent rolling operations.To address the above issues,this study proposes a rapid and accurate obstacle perception method for unmanned roller in construction based on the fusion of millimeter-wave radar and camera.Firstly,the method replaces the convolutional kernel in the feature extraction network of the Faster R-CNN with dilated convolution kernels of different dilation rates to achieve rapid and accurate identification of dam surface obstacles.Subsequently,the distance and velocity information of obstacles perceived by the millimeter-wave radar are fused with the category information perceived by the camera.This fusion achieves comprehensive and accurate perception of the dam surface environment.The Engineering cases indicate that the DC-Faster R-CNN object detection algorithm proposed in this study improves the mAP value by 2.59%compared to the traditional Faster R-CNN object detection algorithm,and reduces the detection time per image by 0.015 s.Additionally,the perception strategy based on multi-modal fusion achieves precise obstacle avoidance during the dam compaction process,enhancing the safety and efficiency of dam compaction construction.