.png){kind=icon}
.png){kind=icon}
Return
A visual perception and distance estimation system based on wavelet-enhanced RT-DETR and SGBM for wheeled-legged quadruped robots in special terrains
.png){kind=icon}
OA
AI
W
R
J
J
X
DOI：10.1016/j.aej.2026.05.048.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Wheeled-legged quadruped robots exhibit strong adaptability to complex terrain environments. However, reliable perception of special terrains such as stairs and slopes remains a critical challenge for autonomous motion decision-making. To address this problem, this study proposes a vision-based special terrain perception and distance estimation system for wheeled-legged quadruped robots. This system develops a wavelet-enhanced real-time detection transformer, known as WLRT-DETR, combined with binocular depth estimation, achieving precise recognition and distance estimation of special terrain targets. First, to achieve the lightweight of the feature extraction network, we design a Re-parameterized Cross-stage Partial Aggregation (ReCSPA) module. By leveraging a dual-path structure combined with channel compression and structural re-parameterization strategies, the module effectively reduces the number of model parameters. Second, to enhance edge representation in feature maps, we introduce a Wavelet-Spatial Fusion Enhancement (WSFE) module, which exploits discrete wavelet transform (DWT) to decompose the frequency domain information of images and deeply fuse it with spatial features, thereby improving detection accuracy. Finally, the candidate boxes generated by WLRT-DETR are used as prior information to constrain the disparity search space, within which the Semi-Global Block Matching (SGBM) algorithm is applied to estimate the spatial distance of the targets, providing reliable depth information for robotic motion decision-making. Extensive experiments were conducted on a self-constructed special terrain dataset to comprehensively evaluate the proposed model by comparing it with state-of-the-art detectors, including representative YOLO-based models and RT-DETR v1/v2. The experimental results demonstrate that the proposed WLRT-DETR achieves superior performance in terms of precision, F1-score, and mean average precision (mAP) while maintaining a lightweight architecture suitable for real-time deployment. Moreover, the stereo distance estimation achieves a relative depth error within 3%, satisfying the practical requirements of wheeled-legged quadruped robots operating in special terrains.
Keywords:
Wheeled-legged quadruped robot
Wavelet-enhanced RT-DETR
Stereo distance estimation
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
6.8
Papers:
6.3K
Citations:
2.6W
Organization
Cited Papers