Edge-Accelerated Underwater Object Detection for AUVs: System Design and Deployment Validation

Abstract

Robust and real-time visual perception is a critical prerequisite for the autonomous operation and closed-loop control of Underwater Vehicles (AUVs/ROVs). However, deploying advanced frequency-domain attention models, such as the Wavelet-based Frequency-Domain Attention (WFDA) network, on power- and compute-constrained underwater edge devices poses significant challenges. Specifically, applying standard INT8 quantization to these models often triggers severe precision collapse, as the coarse quantization steps cause high-frequency responses in Haar wavelet transforms to underflow, obliterating crucial underwater textures. To bridge the gap between complex algorithmic design and practical robotic deployment, this paper proposes an edge-oriented engineering framework. First, we develop a lightweight, multi-threaded interactive monitoring system based on PyQt5, ensuring asynchronous, non-blocking visualization and dynamic parameter configuration for field tests. Second, we introduce a finely-grained mixed-precision quantization strategy using TensorRT on the NVIDIA Jetson Orin Nano platform. By utilizing KL-divergence-based INT8 calibration for standard convolutional layers while enforcing an FP16 fallback mechanism for mathematically sensitive high-frequency nodes, we successfully circumvent numerical underflow. Experimental results demonstrate that our system maintains a high detection accuracy of 90.9% mAP@0.5 while accelerating the inference speed to 32.4 FPS. This performance comfortably exceeds the ¿15 FPS threshold required for stable visual closed-loop control in fluid-damped underwater environments, validating the framework’s capability to deliver edge intelligence for real-world marine robotic applications.