基于遮挡感知像素级融合的单目3D目标检测方法

Monocular 3D Object Detection Based on Occlusion-Aware and Pixel-Level Fusion

利用深度边界框残差和目标边界框来联合估计密集场景深度,对3D目标进行双流检测,从而产生更鲁棒的检测结果。其中,几何流组合了可见深度和深度边界框残差,通过显式的遮挡感知优化方法来恢复目标的3D边界框。此外,基于包围框的几何投影方案被用于增强距离感知。上下文流则用于直接回归3D目标的位置和大小。这种新颖的双流表示促进了跨流之间的一致性,将双流的输出结果进行对齐,从而提高整体性能。在公开数据集上的大量实验表明,该方法在保持实时推理速度的同时,在汽车类别上的检测精度也优于最先进的方法。

This paper proposes to jointly estimate the depth of dense scenes using residual depth bounding boxes and target bounding boxes,and perform dual stream detection on 3D objects to produce more robust detection results.Among them,the geometry stream combines visible depth and depth bounding box residuals,and recovers the 3D bounding box of the target through explicit occlusion-aware optimization methods.In addition,the geometric projection scheme based on bounding boxes is used to enhance distance perception.Context stream is adopted to directly regress the position and size of 3D objects.This novel dual stream representation promotes consistency between cross streams,aligning the output results of the two streams,thereby improving overall performance.Numerous experiments on public datasets have shown that this method outperforms state-of-the-art methods in detecting car class while maintaining real-time inference speed.