Research On 3D Object Detection Algorithm

With the rapid development of deep learning,many artificial intelligence applications have penetrated into people’s daily lives.The automatic driving system is one of them,and the environment perception technology is very important to realize the automatic driving,and the object detection is an important part of the environment perception,so the object detection has great research significance.Lidar has unique advantages in obtaining environmental spatial information.It can obtain high-precision and large-scale spatial depth information and is not susceptible to interference.Therefore,3D object detection based on lidar point cloud is currently a hot research direction.At present,object detection algorithms are mainly divided into object detection based on traditional technology and object detection based on deep learning.Object detection based on deep learning is divided into two types: 2D object detection and 3D object detection.2D object detection algorithms are relatively Mature,the development of 3D Object detection is relatively lagging behind.Faster R-CNN is a more mature and stable object detection algorithm in 2D object detection based on deep learning.However,there are a large number of convolutions and pooling calculations in the Faster R-CNN detection network,which will cause the loss of small target object information.Therefore,the detection accuracy of small target objects is low,and the convolutional network cannot directly convolve the 3D point cloud.Calculation.In response to the above-mentioned problems,the article mainly studies how to perform object detection on 3D point cloud,and how to improve the detection accuracy of small object objects.First of all,in view of the problem that the 3D lidar point cloud cannot be directly extracted using the convolutional neural network,a bird’s eye projection method is proposed to project a highly sparse,discrete and irregular 3D point cloud into a regular 2D point.Clouds bird’s eye view.According to the principle of lidar’s perception of the environment,the lidar point cloud mainly contains the height potential difference information of the surface contour of the object in the space.The point cloud in the grid is projected into pixels corresponding to the 2D overlook image according to the height information to generate a 2D lidar point cloud bird’s-eye view.Secondly,in order to solve the problem of low detection accuracy of the Faster RCNN network for small target objects,the paper proposes a method to improve the feature extraction network structure.The low-level feature map is merged with the high-level feature map through lateral connection to improve The detection network detects the accuracy of small target objects.The experimental data set is completed using the KITTI data set.Experimental results show that the improved structure model not only meets the real-time requirements,but also effectively improves the accuracy of small object object detection.