Infrastructure-based perception systems can significantly contribute to the advancement of Advanced Driver Assistance Systems (ADAS) and connected mobility, thereby enhancing public safety, and enabling early detection of potential hazards and emergencies. The primary objective of this research is to empower AI systems to not only recognize familiar objects but also effectively handle novel, previously unseen data, making them adaptable to the dynamic and unpredictable nature of real-world driving environments.
Key Research Objectives:
- Sensor Data Utilization: Investigate diverse infrastructure sensors to distinguish known and unknown objects, emphasizing those critical for traffic safety.
- Novel Sample Selection and Clustering: Develop strategies to select representative novel samples and perform unsupervised clustering, ensuring efficient handling of unknown objects.
- Incremental Learning: Explore techniques for integrating novel classes into the detection framework while avoiding catastrophic forgetting and adapting to evolving scenarios.
- Domain Knowledge Integration: Incorporate traffic rules and road characteristics into recognition frameworks to enhance environment understanding.
- Evaluation Metrics: Create rigorous evaluation metrics to assess method performance accurately.
Additional research objectives of this Ph.D. thesis include open-set recognition in the field of trajectory analysis and scene understanding tasks. These solutions contribute significantly to identifying unusual patterns in the traffic flow and predicting and optimizing congestion. This Ph.D. project promises to significantly contribute to improving traffic safety and the capabilities of intelligent transportation systems. It tackles the challenges of open-set recognition, learning, and domain knowledge integration within infrastructure sensor data. Ultimately, the research aims to make our roads safer and more adaptive in the face of ever-evolving conditions.
