研究目的
To recognize vehicles in urban environments for Autonomous Land Vehicle navigation using a novel Global Cylindrical Coordination Histogram Descriptor.
研究成果
The proposed GCCHD descriptor, combined with the Adaboost classifier, achieves the best performance in vehicle recognition compared to other global descriptors, with high true positive rates across different distance subgroups. It effectively handles rotation invariance around the z-axis and improves descriptiveness by incorporating azimuth angular information.
研究不足
The performance degrades at longer ranges due to sparser LIDAR points, which may not adequately represent vehicle shapes. The method assumes ground segmentation and object clustering have been performed, and it is primarily tested in urban environments, potentially limiting generalizability to other settings.
1：Experimental Design and Method Selection:
The study involves designing and comparing global descriptors for vehicle recognition from 3D LIDAR point clouds. Methods include Bounding Box descriptor, Histogram of Local Point Level descriptor, Hierarchy descriptor, Spin Image, and the proposed GCCHD. Classifiers used are Support Vector Machine (SVM) and Adaboost, with four-cross-validation to prevent overfitting.
2：Sample Selection and Data Sources:
Two datasets are used: the public Sydney Urban Object dataset (588 objects, 188 vehicles and 400 non-vehicles) and a manually prepared and labeled dataset (1,018,771 samples, 89,651 vehicles and 929,120 non-vehicles). The Sydney dataset was collected with Velodyne LIDAR, and the custom dataset includes subgroups based on distance to the ALV (near: ≤15m, mid: 15-35m, far: >35m).
3：List of Experimental Equipment and Materials:
Velodyne LIDAR is used for data collection. No specific models or brands are detailed beyond this.
4：Experimental Procedures and Operational Workflow:
For each LIDAR scan, ground segmentation and object clustering are performed. Global descriptors are computed for each cluster, and classifiers (SVM and Adaboost) are trained and tested on the datasets. Performance is evaluated using ROC curves and accuracy metrics.
5：Data Analysis Methods:
Statistical analysis includes calculating true positive rates and false positive rates from ROC curves. Parameters for SVM (C=362.0387, γ=0.0884) and Adaboost are specified, with comparisons based on descriptor performance across different distance subgroups.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
