基于点云深度学习的3D目标检测

针对目前利用点云进行3D目标检测的研究较少和检测精度不高的问题,利用Frustum-Pointnets模型实现基于点云的3D目标检测,并在该模型的基础上进行改进,选用不同的激活函数和参数初始化方法进行组合对比,进一步提高模型的精度。实验表明:在选用Swish激活函数和He参数初始化方法时汽车平均检测精度提高了0.31 %,行人平均检测精度提高了0.41 %,骑车人平均检测精度提高了5.5 %。因此改进后的模型能有效提高检测的精度,使得模型能够应用在复杂的场景中。

3D object detection based on point cloud deep learning

In terms of current situation of insufficient researches and accuracy issue for 3D Object Detection of Point Cloud,higher model accuracy is possible to be achieved by improved Frustum-Pointnets model based on 3D Object Detection of Point Cloud,and under the cross-reference by means of different activation and initialization.Experimental results shows that using Swish activation and He initialization for detecting bring about,averagely,0.31% accuracy growth for cars detection,0.41% for pedestrians and 5.5% for cyclist.Accordingly,improved model is able to enhance the detection accuracy to be applied to complex occasions.