On 25 September at 09:15, Shan Wu will defend his doctoral thesis "Deep Learning Architectures for Autonomous Perception: From Sensor Auto-Calibration to Efficient Multi-Object Tracking“ to obtain the degree of Doctor of Philosophy (in Computer Science).
Supervisor:
Assoc. Pro. Amnir Hadachi, University of Tartu
Opponents:
Prof. Pascal Vasseur, Jules Verne University of Picardy (France)
Prof. Haitham Al Hassanieh, Swiss Federal Institute of Technology in Lausanne (Switzerland)
Summary
Self-driving cars and autonomous robots rely on complex perception systems to understand the world around them and operate safely. To do this effectively, they use a combination of different sensors, like standard cameras for rich visual details and LiDAR for precise 3D depth.
However, building these systems presents two major real-world problems that this dissertation addresses:
- Aligning the "Eyes": First, the different sensors must be perfectly aligned with each other, which is traditionally a slow, manual, and labor-intensive process.
- Tracking Moving Objects: Second, once the car can see, it has to track multiple moving objects (termed Multi-Object Tracking, MOT), like pedestrians and other vehicles. The current Machine Learning (ML) models used for this are often too bulky, demanding massive amounts of computer power, and they can easily miss smaller, occluded objects.
This dissertation solves these challenges by creating smarter, more efficient ML models:
- Automatic Calibration: The first part of the thesis introduces "NetCalib," an ML model that acts like an automatic eye-doctor for the vehicle. It learns to iteratively align the camera and LiDAR on its own, completely removing the need for human intervention or specialized physical targets.
- Enhanced Detail Focus: To improve tracking, a Transformer MOT model was upgraded to better focus on small, local details, making it much safer and more accurate at detecting smaller or distant pedestrians.
- "Green" and Efficient AI: Finally, the research slims down the MOT model into a "green," highly efficient architecture called MOTT. By removing unnecessary components, this new model runs almost twice as fast and uses significantly less computer power, all while maintaining top-tier accuracy.
Ultimately, this work makes the digital brains of autonomous vehicles faster, significantly more energy-efficient, and better prepared for real-world roads.
