Real-Time 3D Multi-Object Detection and Localization Based on Deep Learning for Road and Railway Smart Mobility

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Environmental perception is crucial for smart mobility, autonomous vehicles, and advanced driver-assistance systems (ADASs).
• Accurate perception enables enhanced decision-making and high-precision actions in complex traffic scenarios.
• Existing methods often lack real-time performance or applicability across diverse transportation modes.

Purpose of the Study:

• To introduce a novel real-time deep learning approach for 3D multi-object detection.
• To extend environmental perception capabilities for smart mobility applications on both roads and railways.
• To improve the accuracy and efficiency of 3D object localization, dimension, and orientation estimation.

Main Methods:

• Modified the YOLOv3 (You Only Look Once version 3) 2D object detector to predict 3D bounding box parameters.
• Incorporated prediction of 3D object localization, dimensions, and orientation.
• Evaluated the method on the KITTI road dataset and a custom hybrid virtual road/rail dataset from Grand Theft Auto V.

Main Results:

• Achieved good accuracy in 3D multi-object detection on both road and rail datasets.
• Demonstrated real-time performance suitable for operational road and rail traffic environments.
• Highlighted the critical impact of accurate region of interest (RoI) prediction on 3D bounding box parameter estimation.

Conclusions:

• The proposed real-time deep learning method effectively performs 3D multi-object detection for smart mobility.
• The approach is versatile, applicable to both road and rail environments, enhancing situational awareness.
• Accurate RoI prediction is vital for robust 3D object detection in autonomous systems.