Recently, the research team from the Hefei Institutes of Physical Science of the Chinese Academy of Sciences, proposed a point cloud 3D object detection method based on attention mechanism and data augmentation.
“It can help self-driving cars better detect small objects,” said Prof. WANG Zhiling, who led the team.
The achievement has been accepted by IEEE Transactions on Intelligent Transportation Systems and published online as "Early Access."
3D object detection is crucial for autonomous vehicles. It utilizes point cloud data generated by LiDAR to help autonomous vehicles identify surrounding objects. This technology is essential for the safety and efficiency of autonomous driving. Traditional methods usually convert sparse and unordered point cloud data into pseudo-images to extract ordered information. However, this conversion often loses critical features, leading to a decline in detection accuracy, especially in detecting smaller objects.
In this study, researchers introduce a new approach to 3D object detection, SCNet3D. It focuses on improving feature enhancement, preserving information, and detecting small objects by focusing on both feature and data aspects.
With this method, a Feature Enhancement Module is used, which applies an attention mechanism to focus on important features across three dimensions, gradually improving the 3D features from local to global. Also, STMod-Convolution Network (SCNet), which has two channels for feature extraction, is included. One channel works on basic features, while the other handles more complex, advanced features by combining information from bird-eye view pseudo-images.
In addition, the research proposes a Shape and Distance Aware Data Augmentation method, which adds useful samples to the point cloud during training.
Tests proved that this method has many advantages in detecting small objects, even in challenging environments with lots of interference. This makes it a promising tool for autonomous driving.
Overview of SCNet3D （Image by WANG Zhiling）
