用于钢铁回收中的目标检测与尺寸估计算法

为解决钢铁回收过程中由于开放式复杂场景而无法准确地获得目标废钢点云进行尺寸估计,文章提出一种基于Mask R-CNN模型预测掩膜自适应颜色阈值的目标点云提取算法。针对Mask R-CNN模型边缘分割的不完整性,通过预测框对点云进行截取,利用自适应颜色滤波阈值对截取的点云进行HSV颜色空间滤波,将非目标点云离散化,最后通过欧式聚类对目标点云进行分割并进行尺寸估计。通过废旧汽车B柱模拟回收场景,获取汽车B柱点云,验证了该算法的有效性。

To solve the problem of inability to accurately obtain the target scrap point cloud for size estimation due to open and complex scenes in the process of steel recycling, a target point cloud extraction algorithm based on the prediction mask generated by the Mask R-CNN model and adaptive color thresholding technique is proposed. For the incomplete edge segmentation of the Mask R-CNN model, the point cloud is intercepted through the prediction frame, the adaptive color filter threshold is used to filter the intercepted point cloud in HSV color space to discretize the non-target point cloud, and finally the target point cloud is segmented through European clustering and dimensionally estimated. The effectiveness of the algorithm is verified by simulating the recovery scene of the B-pillar of the scrapped car and obtaining the point cloud of the B-pillar of the car.