Efficient Object Detection Using Semantic Region of Interest Generation with Light-Weighted LiDAR Clustering in...

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Convolutional Neural Networks (CNNs) are widely used for object detection in 3D data.
• 3D data-based algorithms offer stability but demand significant computational resources, limiting their use in embedded systems.
• Existing 2D CNNs struggle with lighting variations and require complex processing for 3D data.

Purpose of the Study:

• To develop a computationally efficient method for 3D object detection on lightweight embedded systems.
• To enhance the accuracy of 2D CNN object detection by integrating 3D data characteristics.
• To overcome the limitations of traditional 3D and 2D object detection methods in varying environmental conditions.

Main Methods:

• Preprocessing LiDAR point cloud data and employing clustering for object separation.
• Extracting physical characteristics from 3D data for semantic detection using machine learning classifiers.
• Generating 2D image regions from detected 3D objects to perform 2D CNN object detection, bypassing bounding box tracking.

Main Results:

• The proposed model achieved 81.84% accuracy using YOLO v5 on an embedded board, surpassing typical models by 1.92%.
• Demonstrated superior performance in non-optimal lighting conditions, with 47.41% accuracy in high brightness (+40%) and 54.12% in low brightness (-40%).
• Achieved significant accuracy improvements (8.97% and 13.58% respectively) over general models in challenging lighting.

Conclusions:

• The proposed method effectively utilizes 3D data characteristics to enhance 2D CNN object detection accuracy in resource-constrained environments.
• The approach offers a lighter system alternative to purely 3D data-based detection algorithms.
• The technique shows robustness in varying lighting conditions and reduces execution time, making it suitable for real-world embedded applications.