Design and Implementation of an Adaptive Visual Perception Verification Platform for Amphibious Rescue Robots

Abstract

Amphibious rescue robots encounter significant environmental variations during cross-medium rescue missions, making it challenging for a single visual detection model to simultaneously meet the high-precision requirements of both aquatic and terrestrial scenes. To address this issue, this paper designs and implements an algorithm verification platform for adaptive visual perception, aiming to validate the feasibility of an “environment perception–model scheduling” adaptive pipeline. The platform adopts a modular and hierarchical architecture, integrating a rapid environment discrimination module based on the YOLO framework and two specialized object detection models. Through real-time scene judgment and dynamic model scheduling, it achieves environment-adaptive detection. In addition, the platform provides a graphical user interface and multi-dimensional performance monitoring, supporting multiple input sources such as cameras, videos, and images, thereby supporting algorithm verification in laboratory environments. Experimental results demonstrate that the platform achieves an environment classification accuracy of 92.3%, an average processing frame rate of 28.6 FPS, and a comprehensive detection accuracy improvement of 7.2% in mixed environments compared to the best-performing single model. The evaluation was conducted on a self-constructed dataset comprising 8,000 images and a 2,000-frame mixed-environment video sequence. These results effectively validate the engineering feasibility of the adaptive scheduling mechanism in terms of both accuracy and real-time performance. This platform offers a reusable verification foundation for the subsequent embedded integration and field deployment of amphibious robot visual perception systems.