Enriching 3D Object Detection in Autonomous Driving for Emergency Scenarios: Leveraging Point Cloud Data with CARLA Simulator for Automated Annotation of Rare 3D Objects

Engineering notes

Our investigation demonstrates the superior performance of models trained on a fusion of real-world 3D point cloud data and computer-generated data from CARLA compared to models trained solely on restricted examples.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Advances in 3D object detection for autonomous driving primarily target identifying common entities like automobiles and pedestrians. However, the heavy reliance on extensive, precisely marked training data restricts the ability to identify specialized and uncommon items such as law enforcement vehicles and emergency medical service transports. To address this limitation, this research introduces an inventive methodology utilizing the CARLA simulator to autonomously mark these infrequent objects within 3D point cloud data, enhancing datasets with limited labels. Our investigation demonstrates the superior performance of models trained on a fusion of real-world 3D point cloud data and computer-generated data from CARLA compared to models trained solely on restricted examples. Bridging the gap between simulated and real-world scenarios is achieved by employing diverse simulations in CARLA, covering various weather conditions, human figures, and environmental settings. Training detectors using CARLA-generated 3D point cloud data spanning diverse object classes display substantial enhancements, validated against available datasets for public scrutiny. Additionally, Access to synthetic datasets focusing on rare objects like law enforcement vehicles and medical transports is provided, along with detailed guidelines for creating custom datasets in CARLA, aimed at enhancing data for critical, infrequent items in emergencies.

Links and sources

• Official / arXiv page

Comments