Communication-Efficient Object Detection for Vehicular Applications

Sammanfattning

The increasing adoption of edge devices in vehicular applications underscores the critical role of real-time object detection in enhancing vehicular safety and efficiency. Traditional centralized training approaches for object detection models pose significant challenges, particularly regarding data privacy, communication latency, and scalability in distributed vehicular networks. Federated Learning (FL) offers a promising alternative by enabling collaborative training without requiring data centralization, preserving privacy, and enhancing security. This thesis investigates the integration of model compression techniques, specifically Quantized Stochastic Gradient Descent (QSGD), within the FL framework for vehicular object detection tasks. By focusing on YOLOv8, a state-of-the-art real-time object detection model, this research explores the impact of QSGD on communication efficiency, detection accuracy, and overall model performance in scenarios characterized by non-independent and identically distributed (Non-IID) data. Comprehensive experiments were conducted on automotive datasets, including the KITTI and Zenseact Open Dataset (ZOD), to evaluate the effectiveness of the proposed approach. The results demonstrate that integrating QSGD significantly reduces communication overhead while maintaining competitive detection accuracy, even under challenging data distribution scenarios. This research highlights the feasibility of combining federated learning and model compression, offering valuable insights for developing scalable, communication-efficient, and privacy-preserving solutions in the future of vehicular technology.