LiDAR Dynamic Target Detection Based on Multidimensional Features.

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Sensor Fusion

Background:

• Traditional LiDAR dynamic target detection methods often rely on heuristic thresholding or supplementary data, limiting their robustness and efficiency.
• Existing approaches struggle with accurate point cloud cluster pairing across adjacent frames, hindering reliable motion state evaluation.

Purpose of the Study:

• To develop an innovative LiDAR dynamic target detection method that overcomes the limitations of existing techniques by utilizing multidimensional features.
• To enhance the precision and efficiency of dynamic target detection in LiDAR data through advanced algorithms for point cloud registration and classification.

Main Methods:

• A double registration algorithm (ICP, RANSAC) for precise point cloud cluster pairing between adjacent frames.
• Development of a classification feature system using XGBoost, optimized by a Spearman's rank correlation coefficient-bidirectional search for dimensionality reduction.
• A double Boyer-Moore voting-sliding window algorithm for refining detection accuracy after initial XGBoost classification.

Main Results:

• Achieved a 92.41% correct detection rate for dynamic LiDAR targets.
• Maintained a low static target error detection rate of 1.43%.
• Demonstrated high detection efficiency with a processing time of 0.0299 seconds per frame.

Conclusions:

• The proposed multidimensional feature-based method significantly improves LiDAR dynamic target detection accuracy and efficiency.
• The novel approach offers a robust and automated solution, outperforming existing open-source comparative methods.
• This work provides a valuable advancement for autonomous systems requiring reliable real-time dynamic object recognition.