Federated Object Detection Scenarios for Intelligent Vehicles: Review, Case Studies, Experiments and Discussions

The performance of intelligent vehicles (IVs) in object detection relies not only on the design or scale of the CNN model they use but also on how effectively they share their acquired knowledge with others. Federated learning (FL) is regarded as a unique platform for running distributed and collaborative object detection scenarios, enabling a group of vehicles to distribute their object detection expertise to other groups. Object detectors are the critical software component of intelligent vehicles. Each intelligent vehicle should be able to detect the objects captured by its on-board camera. A FL allows a group of IVsto collaboratively train their detectors andlets them achieve a decent performance even with limited training data. In this study, we provide a survey and experimental analysis of federated object detection (FOD) between IVs and address several existing challengesin federated detection. Furthermore, we examine object detection and training performance with state-of-the-art model aggregation methods. We explore novel and advanced approaches that may enhance collective detection performance. As there is currently no comprehensive review available in the existing literature on FOD with IVs, we reviewed a few existing methods and platforms. We run experiments on multiple well-known public autonomous driving datasets and investigate methods to enhance FOD performance across various data distribution cases