基于2D先验的3D目标判定算法

提出一种基于2D先验的3D目标判定算法.首先用轻量级MobileNet网络替换经典SSD的VGG-16网络,构建出MobileNet-SSD目标检测模型;其次,通过改进网络结构,提高模型对小目标的检测能力,并引入Focal Loss函数来解决正负样本不均衡和易分样本占比较高的问题;在相同数据集上,将改进算法与Faster R-CNN、 YOLOv3及MobileNet-SSD进行对比测试,其平均精度mAP分别提高了7.2%、 8.8%和10.6%;最后,通过改进算法获取ROI,利用深度相机将二维ROI转换为ROI点云,并借助直通滤波来判断目标物体是否为真实场景物体,既省去了传统点云识别中的诸多步骤又避免了点云深度学习中三维数据集制作难度较大的问题,在识别速度和识别精度上达到了较好的平衡.

A 3D object discrimination algorithm based on 2D prior

A 3D object recognition algorithm based on 2D priori is proposed to address the problems of 2D object detection can"t distinguish non-real scene objects from real scene objects and long work cycle of using 3D object recognition directly. Firstly, the VGG-16 network of classical SSD is replaced by the lightweight MobileNet network, and the depthwise separable convolution is used to reduce the model size and construct the MobileNet-SSD object detection model; Secondly, by improving the network structure, two additional prediction feature layers are added to improve the detection ability of the model for small and medium-sized objects, and the Focal Loss function is introduced to solve the problems of unbalanced positive and negative samples and high percentage of easy samples. The improved algorithm was tested against Faster R-CNN, YOLOv3 and MobileNet-SSD on the same dataset, and its mAP improved by 7.2%, 8.8% and 10.6%, respectively. Finally, the ROI is obtained by the improved algorithm, and the two-dimensional ROI is converted into ROI point cloud by the depth camera, and the ROI point cloud is filtered by the PassThrough filter, and whether it is a real scene object is judged according to the filtering result, which eliminates the steps of point cloud preprocessing, key points extraction, feature representations, feature matching in the traditional point cloud recognition and avoids the difficulty of making 3D dataset in point cloud deep learning, achieves a good balance in recognition speed and recognition accuracy.