Baidu Intelligent Cloud
Since 2016, the ASTRI Smart Mobility Technologies (SMT) team has been dedicated to research and development of the smart mobility plan for Hong Kong. Conventionally, smart mobility solutions rely on rule-based methods or end-to-end task-specific neural networks for traffic data processing. The endless long-tail problems inherent with such methods have hindered commercial deployment of roadside and onboard perception and decision technologies. Inspired by recent development of large pre-trained AI models, we propose to develop large pre-trained models with multi-modal traffic data as the base for next-generation smart mobility solutions. In this project, ASTRI SMT team will tackle four fundamental factors for enabling large traffic model pre-training: - Construction of large-scale multi-modal traffic dataset and validation of data influence. - Scalable model design from millions to billions of parameters for multi-modal traffic data input. - Design of self-supervised learning task for large model pre-training. - Progressive feasibility plan with constrained infrastructure support. The pre-trained large traffic model could benefit numerous downstream smart mobility tasks, such as anomaly detection, driver's behavior analysis, data-driven traffic simulation, and so on. With general knowledge encoded into the large base model, downstream task-specific models will potentially be able to handle rare and unseen cases, and reduce the risks of accidents or congestion on road. Technologies enabled by this project is key to the Hong Kong smart mobility roadmap.