DefinePK

Abstract

The proliferation of intelligent transport systems demands sophisticated tools to understand and optimize driver behavior without compromising privacy or system integration. This research proposes a groundbreaking framework leveraging zero-permission embedded sensors, such as accelerometers, gyroscopes, and magnetometers, in mobile and in-vehicle devices to profile driver behavior comprehensively. The core innovation lies in employing AI-enhanced edge analytics powered by federated learning, ensuring data processing remains local to the device. This approach addresses privacy concerns while maintaining high model accuracy. By integrating self-supervised learning (SSL) techniques with sensor data streams, the system autonomously detects and labels complex driving patterns such as aggressive acceleration, harsh braking, and distracted driving. Additionally, the framework utilizes Graph Neural Networks (GNNs) to model and analyze dynamic road environments and driver-vehicle interactions, offering unparalleled insights into driving behavior in diverse traffic conditions. Generative AI models, such as diffusion-based architectures, simulate potential driving scenarios, aiding real-time decision support systems. This research showcases exceptional scalability by deploying energy-efficient AI accelerators for on-device inference, enabling continuous monitoring with minimal power consumption. Experimental validation on datasets from smart cities demonstrates enhanced behavioral profiling accuracy (98%) and latency reductions (40%) compared to traditional cloud-based models. Applications include advanced driver-assistance systems (ADAS), fleet management optimization, and insurance telematics, paving the way for safer, more efficient, and driver-centric transport ecosystems.