研究目的
To accurately distinguish pedestrians and vehicles using machine learning from 3D point cloud data acquired by a 3D lidar, specifically recognizing cars, bicyclists, and pedestrians with improved accuracy and lower processing time.
研究成果
The proposed SVM-based method with eight-dimensional features provides higher recognition accuracy (95.5%) and lower processing time compared to random forest and 26-dimensional feature methods, making it suitable for low computational environments. Future work involves embedding the method into a tracking system and testing in varied environments.
研究不足
The data set does not include lidar data related to buses, trucks, and various pedestrians (e.g., with umbrellas or pushing baby buggies), limiting generalization. The method may not perform well in all road environments, and further evaluation in diverse scenarios is needed.
1：Experimental Design and Method Selection:
The study uses a support vector machine (SVM) classifier with an RBF kernel for multiclass classification (cars, bicyclists, pedestrians) based on eight-dimensional features extracted from 3D lidar data. The one-vs-one method is employed for multiclass SVM.
2：Sample Selection and Data Sources:
The Stanford Track Collection data set is used, containing lidar measurements of cars, bicyclists, and pedestrians within 70 m of the lidar. Training data includes 32,346 scans, and test data includes 16,846 scans.
3：List of Experimental Equipment and Materials:
A Velodyne HDL-64ES2 64-layer lidar is used for data acquisition. A computer with an Intel Core i7-6700K CPU, 32 GB RAM, Windows 10 OS, and C++ programming language is used for processing.
4：Experimental Procedures and Operational Workflow:
Lidar data is clustered using background subtraction, features are extracted (e.g., number of points, distance, reflection intensity, size ratios, velocity), and SVM classification is performed. Hyperparameters (C and γ) are optimized using grid search and cross-validation.
5：Data Analysis Methods:
Recognition accuracy, precision, recall, F-measure, and processing time are evaluated and compared with random forest and a 26-dimensional feature method.
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
