A dataset for investigating dust removal for person detection

Autonomous agricultural machinery must be able to operate safely in harsh conditions, such as dust. Dust raised during tillage can obscure objects such as an approaching person, thus posing a safety risk. One way to cope with this is to remove the dust from the image data, thereby increasing the detectability of the person. Hence, this study investigates the extent to which dust removal improves person detection. A dataset containing real-world dusty and dust-free images is presented and enriched with synthetic dusty images. To achieve this, dust was added synthetically to the dust-free images. The developed method recreates the temporal formation of dust during tillage and its influence by wind. Subsequently, state-of-the-art image dehazing and translation methods were benchmarked on the dataset. Additionally, a YOLO11 model was utilized for person detection and trained on different dataset configurations. Training on synthetic dusty images increased detection performance, measured in terms of mAP, by 55.11 percentage points compared to a pretrained detection network. Furthermore, combining synthetic and real-world dusty images increased mAP by 2.78 percentage points compared to training on real-world dusty images alone. Moreover, dust removal improved the detection performance in heavy dusty conditions by 27.23 percentage points compared to a pretrained network. More importantly, the combination of a dust removal method and a retrained YOLO11 on the specific output of the dust removal method outperformed a YOLO11 trained on real-world dusty images by 3.39 percentage points in heavy dusty conditions. The results highlight the potential to enhance the safety of autonomous agricultural machinery by removing dust from images prior to detection.