Assistive navigation for visually impaired users requires reliable obstacle detection on portable, low-cost hardware. We present AI Navigo, a wearable obstacle-awareness system that combines monocular video, YOLO-based object detection, and audio feedback delivered through Bluetooth earbuds. The system is implemented on a Raspberry Pi with an endoscope camera mounted on eyeglasses and it is designed to detect four outdoor obstacle classes relevant to pedestrian safety: (a) cars, (b) persons, (c) traffic cones, and (d) potholes. We evaluate YOLOv9 and YOLOv10 on a dataset of 6,276 images collected from outdoor pedestrian environments and compare their performance under multiple learning-rate and epoch settings. YOLOv9 achieved the strongest overall results, reaching a detection score of 0.847 and outperforming YOLOv10 in difficult classes such as potholes and persons, while also showing more stable validation behavior and better generalization. The final deployment pipeline converts detections into spatial audio descriptions using a lightweight coordinate-to-description mapping function. Although inference speed on the Raspberry Pi remains a practical limitation, the results demonstrate the feasibility of low-cost vision-based obstacle awareness for wearable navigation assistance. Future work will focus on improving runtime efficiency, expanding obstacle coverage, and evaluating the system in real user settings.
Asvini Thivakaran, Vani Kandasamy· Annual International Compute...· 0 citations
Identifying proteins that neutralize snake venom toxins is a critical bottleneck in antivenom development, constrained by the scarcity of experimentally resolved toxin-binder structures and the high cost of wet-lab screening. This paper presents a computational screening pipeline that prioritizes toxinbinder candidates for downstream experimental validation, addressing the challenge of candidate ranking when labeled data is limited and supervised models risk overfitting. A graph reranking algorithm, BinderGraph, is proposed: it propagates frozen ESM2 cosine similarity scores across a binder co-occurrence graph constructed from structural training data, without any learned parameters. Evaluated across 10 random toxin-level splits on 61 unique toxins, Frozen ESM-8M with BinderGraph achieves mean Recall@5 of $0.836 \pm 0.077$, outperforming a trained MLP DuaIEncoder (Recall@5 = 0.804 ± 0.099) and a larger ESM2-650M model. These results demonstrate that biologically informed post-processing of pretrained embeddings is more effective than additional parameters or model scale when training data is scarce. Intended as a first-stage filter for experimental followup such as surface plasmon resonance (SPR) and enzyme-linked immunosorbent assays (ELISA), all comparisons are subjected to multi-split statistical significance testing.
