3D Objective Detection for Autonomous Driving based on Two-stage Approach

Abstract

As one of the hottest areas in the current technology industry, the field of autonomous driving has attracted the attention of many technology workers. How to use point cloud data for accurate multi-objective prediction is a key issue, which includes 3D object detection and multi-object tracking. CenterPoint proposes a novel anchor-free, two-stage 3D object detection method. The first stage uses a CenterNet approach, that is, using the center point to represent the object, using the feature map after feature extraction as input, and outputting a heatmap of the probability of the location of the center of the object for each category to predict the location of the target object, and obtaining other properties from the feature regression of the point location. The second stage is to extract features from the center point of the bounding box of the prediction target to refine the prediction results. However, the 3D backbone network of the CenterPoint has the disadvantages of low feature extraction accuracy and low second stage refinement accuracy. In order to solve these problems, this paper proposes to use VoxelResBackBone8x based on deep residual network Resnet as the 3D backbone network, simplify the 2D backbone network to improve feature extraction accuracy, and use the Set Abstraction Module to make the model use both the processed advanced features and the original point cloud features to further improve the accuracy.