Mobile river‑monitoring robots must perceive obstacles and water boundaries that cannot be fully described by geographic waypoints alone. On resource‑constrained platforms, turning imperfect visual predictions into timely and inspectable guidance is a major challenge. Conventional object labels or steering commands lack explanations of the supporting evidence and cannot indicate when that evidence is unreliable. To address this, we introduce PAANI, an on‑device perception‑to‑guidance architecture.\
PAANI couples a project‑trained YOLO11n detector with a custom MobileNetV3‑Small semantic segmenter, and performs timestamp‑aligned evidence fusion on an Arduino UNO Q.\
- Detector: Trained on 10,000 WaterScenes images for four‑class detection; the FP32 ONNX model occupies 14.817 MB, achieving mAP@0.5IoU of 0.7388 (checkpoint) and 0.7367 after ONNX export validation.\
- Segmenter: Trained on 1,127 MaSTr1325 images, yielding a validation mIoU of 0.9750.\
- Bounded Tracking: Supplies object persistence across frames, ensuring consistent identity.\
- Corridor Policy: An explicit policy that merges surface labels, accepted detections, urgency levels, and mask uncertainty to produce the final advisory. Each advisory reveals its contributing evidence and policy rationale, achieving explainability.\
ROS 2 interfaces connect the local AI pipeline to a separate Gazebo vessel, localization, and control testbed, enabling end‑to‑end simulation.\
In a five‑minute UNO Q recording, the pipeline exhibited a median latency of 467.8 ms and a 95th‑percentile latency of 580.3 ms at a configured 0.5 Hz cadence. Evaluation uncovered black‑input misclassification and a sampling‑rate mismatch that prevented the diagnostic apparent‑motion estimator from gathering sufficient evidence.\
These results demonstrate that PAANI provides an inspectable and reusable edge‑robotics foundation while clearly separating model accuracy, on‑board execution performance, and validated on‑water collision‑avoidance capability.\
Review: PAANI achieves high‑precision detection and segmentation fusion on a severely constrained embedded platform, and its corridor policy delivers explainable navigation guidance, offering a solid technical pathway for real‑world river‑robot deployments.