엘레베이터 버튼 검출성능 개선을 위한 어텐션 기반의 DETR 모델 개발
Development of Attention-based DETR Model for Improving Elevator Button Detection Accuracy
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Delivery robots operating in multi-story buildings must not only navigate flat surfaces but also move freely between floors, which requires reliable elevator button recognition technique. Although many deep learning models have been proposed for object detection, elevator button detection remains challenges due to variable lighting, reflective surfaces, and visual similarity to surrounding elements. To address the problems, we propose an enhanced object detection architecture that integrates an Attention Module into DETR, a representative end-to-end detection framework. In particular, we introduce the Deformable Convolution Attention Module (DCAM), which leverages deformable convolutions to improve spatial adaptability and object recognition accuracy. When DCAM is applied to the existing DAB-DETR model, the proposed method achieved a 3.61% improvement in mean Average Precision (mAP IoU(50:95)) and an 8.66% increase in Average Recall (AR maxDets(100)) compared to the baseline, demonstrating the effectiveness of our approach in complex visual environments.
Abstract
Delivery robots operating in multi-story buildings must not only navigate flat surfaces but also move freely between floors, which requires reliable elevator button recognition technique. Although many deep learning models have been proposed for object detection, elevator button detection remains challenges due to variable lighting, reflective surfaces, and visual similarity to surrounding elements. To address the problems, we propose an enhanced object detection architecture that integrates an Attention Module into DETR, a representative end-to-end detection framework. In particular, we introduce the Deformable Convolution Attention Module (DCAM), which leverages deformable convolutions to improve spatial adaptability and object recognition accuracy. When DCAM is applied to the existing DAB-DETR model, the proposed method achieved a 3.61% improvement in mean Average Precision (mAP IoU(50:95)) and an 8.66% increase in Average Recall (AR maxDets(100)) compared to the baseline, demonstrating the effectiveness of our approach in complex visual environments.