Comparison of Pedestrian Detectors for LiDAR Sensor Trained on Custom Synthetic, Real and Mixed Datasets

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous vehicles rely on deep learning for object detection, necessitating large labeled datasets.
• Current data collection and labeling are time-consuming and often sensor-specific.

Purpose of the Study:

• To investigate the effectiveness of using synthetically generated and mixed datasets for training LiDAR pedestrian detection algorithms.
• To evaluate if synthetic data can enhance the performance of deep learning models in autonomous driving.

Main Methods:

• Simulated road environments using the 3D rendering Carla engine to generate synthetic LiDAR data.
• Developed an automated pipeline for labeling and reshaping synthetic data into range images for deep learning.
• Trained and validated the YOLOv4 neural network architecture using synthetic, real (Waymo open dataset), and mixed datasets.

Main Results:

• The YOLOv4 model trained on a mixed dataset showed improved precision and recall compared to models trained solely on real or synthetic data.
• Achieved an F1-score of 0.84, indicating robust pedestrian detection performance.
• Demonstrated the viability of synthetic data generation for improving LiDAR-based object detection.

Conclusions:

• Synthetically generated data, when combined with real data, can significantly enhance the performance of LiDAR pedestrian detection models.
• The proposed approach offers a more efficient method for data acquisition and model training in autonomous vehicle development.