This study investigates the optimization of Light Detection and Ranging (LiDAR) object detection on roadside data for automated driving through transfer-learning. A State-of-the-art object detection algorithm, trained on an extensive dataset from Automated Vehicles (AVs), is evaluated for its applicability to roadside sensors. To address the scarcity of roadside data, a novel sensor setup was introduced, enabling simultaneous data collection from AVs and roadside poles. The performance of a pre-trained model on both AV and roadside data demonstrated a significant decline, with the roadside data yielding only 82% of the performance achieved on AV data. However, by fine-tuning the pre-trained model with a small subset of roadside data, a remarkable improvement to 94% of the AV data based performance was achieved. These findings highlight the necessity of transfer-learning and emphasize its potential to enhance object detection accuracy in roadside scenarios. This study contributes to the advancement of automated driving technology, paving the way for more safe and efficient transportation systems.
Transfer-Learning of a LiDAR Detection Algorithm Pre-Trained on Autonomous Driving Data for Roadside Infrastructure Application
24.09.2023
8228784 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
British Library Conference Proceedings | 2020
|