Real-Time Detection of Non-Stationary Objects Using Intensity Data in Automotive LiDAR SLAM.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Accurate mapping and localization are crucial for autonomous vehicles.
• Distinguishing static environments from dynamic objects is a key challenge in Simultaneous Localization and Mapping (SLAM).
• Traditional methods struggle with non-stationary objects like vehicles, pedestrians, trees, and lawns in 3-D LiDAR data.

Purpose of the Study:

• To demonstrate the feasibility of deep learning for real-time detection of non-stationary objects in 3-D LiDAR point clouds.
• To develop a method that alleviates the need for large, point-wise annotated 3-D datasets.
• To improve the performance of laser-based SLAM systems by accurately segmenting motion.

Main Methods:

• Utilized deep learning models trained on 2-D grayscale images synthesized from LiDAR intensity data.
• Employed both supervised and unsupervised training processes for neural networks.
• Filtered 3-D point clouds using labeled pixels from corresponding 2-D intensity images.

Main Results:

• Successfully detected non-stationary objects using neural networks trained on 2-D images.
• Demonstrated that this approach mitigates the scarcity of 3-D annotated datasets.
• Showcased improved localization accuracy and map consistency in a laser-based SLAM system.

Conclusions:

• Modern deep learning techniques are feasible for real-time non-stationary object detection in 3-D LiDAR data.
• Synthesizing 2-D images from LiDAR intensity data offers a viable alternative for training detection models.
• The proposed method enhances the robustness and reliability of automotive SLAM systems.