A hybrid computer-vision pipeline for detecting walking hazards on sidewalks. Its novel contribution is chaining two unsupervised anomaly detectors: a Variational Autoencoder first flags out-of-distribution frames by their high reconstruction error, and a One-Class Support Vector Machine then classifies whether a flagged anomaly is genuinely hazardous, suppressing false alarms from benign anomalies such as manhole or water-valve covers. Trained and tested on recorded sidewalk video, the VAE achieves an AUC of 0.94 and the combined system reaches 91.4% accuracy, giving a reliable hazard-detection method for safe outdoor navigation.
Presents a computer-vision system for detecting walking hazards on sidewalks to support safe navigation in unpredictable outdoor settings, using a two-stage hybrid of a Variational Autoencoder (VAE) …