Combined Edge- and Stixel-based Object Detection in 3D Point Cloud.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Environment perception is crucial for autonomous vehicle path planning and safety.
• Raw sensor data from cameras, LiDAR, and IMUs lack object identification, hindering safe operation.
• Existing methods struggle with precise 3D point cloud segmentation for vital object detection.

Purpose of the Study:

• To propose an improved edge-oriented segmentation method for accurate 3D object detection from point clouds.
• To enhance the precision and efficiency of object segmentation for autonomous driving systems.
• To facilitate visualization in advanced driving assistance systems (ADAS) through reduced computational time and data points.

Main Methods:

• Object bounding areas identified using edge detection and stixel estimation on 2D stereo camera images.
• Reconstruction of 3D sparse point clouds for identified objects within bounding areas.
• Segmentation of dense point clouds by matching reconstructed 3D sparse object point clouds with the scene's point cloud.

Main Results:

• The proposed edge-oriented segmentation method demonstrates improved precision in 3D point cloud segmentation.
• Experimental results confirm accurate segmentation of objects from the 3D point cloud.
• Significant reduction in computational time and the number of points in object point clouds observed.

Conclusions:

• The improved edge-oriented segmentation method effectively enhances 3D object detection accuracy for autonomous vehicles.
• Reduced computational demands and data size facilitate easier integration and visualization in ADAS.
• This approach contributes to safer and more efficient autonomous driving operations through reliable perception.