Conception of a High-Level Perception and Localization System for Autonomous Driving

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Sensor Fusion

Background:

• Autonomous driving systems require robust perception and localization, especially in GPS-denied environments.
• Existing systems often face challenges with scalability, autonomy duration, and real-time processing.
• Development of integrated, multi-modal sensor platforms is crucial for advancing autonomous vehicle capabilities.

Purpose of the Study:

• To design and implement a high-level, compact, scalable, and long-autonomy perception and localization system for autonomous driving.
• To create a multi-modal dataset for training and evaluating perception and localization algorithms.
• To facilitate research and development in autonomous vehicle planning and control modules.

Main Methods:

• Utilized a benchmark system comprising a 128-channel high-resolution lidar, stereo global shutter camera, inertial navigation system, time server, and embedded computer.
• Integrated Recurrent Bayesian Neural Network (RBNN) detection, Deep Convolutional Neural Network (DCNN) detection, and lidar-based localization algorithms.
• Collected and annotated a dataset of 10,000 lidar frames under diverse weather and traffic conditions for algorithm training and evaluation.

Main Results:

• Developed a perception and localization system capable of real-time object detection and localization in challenging, GPS-denied environments.
• Achieved performance competitive with state-of-the-art algorithms, with processing times suitable for real-time autonomous driving applications.
• Successfully built and utilized a multi-modal dataset to train and validate the embedded perception and localization algorithms.

Conclusions:

• The conceived system offers a compact, scalable, and long-autonomy solution for autonomous driving perception and localization.
• The system's real-time capabilities and accurate advanced outputs significantly aid researchers and engineers in developing planning and control modules.
• This work contributes to democratizing access to autonomous vehicle research platforms by providing a comprehensive and effective system.