An Empirical Exploration of Unsupervised Mobile Object Detection in a V2I Scenario
Loading...
Date
2025-05
Authors
Kiener, Jonathan
Journal Title
Journal ISSN
Volume Title
Publisher
The Ohio State University
Abstract
As object detection models for self-driving vehicles continue to advance, the scarcity of labeled training data intensifies as a bottleneck to the development process. While some methods for unsupervised learning exist, none have demonstrated sufficient performance on a vast and diverse testing set to warrant a widespread shift in approach. In this paper, we introduce UDAT (Unsupervised Detectors for Annotation and Training), a method by which roadside units (RSUs) can learn to detect mobile objects, creating labels that could be used to train vehicle detectors. The immobility of the units provides a consistency that allows for much more accurate predictions than unsupervised vehicle-mounted systems. Our approach closely mirrors the state-of-the-art MODEST[37] process, using the clustering of ephemeral points across LiDAR frames to create bounding boxes, which are then used as pseudo-labels to train an off-the-shelf detector. Multiple rounds of self-training increase the accuracy of the output annotations. We aim to demonstrate that this training regimen performs successfully enough to justify real-world viability as a roadside data annotator for the training of deep-learning vehicle detectors.
Description
Keywords
Self Driving, Unsupervised Learning, Autonomous Vehicles, Roadside Detection, LiDAR Perception, Artificial Intelligence, Machine Learning