Detecting Hazardous Materials in Multispectral Images

Abstract

The increasing deployment of autonomous vehicles in various scenarios offers the advantage of protecting human lives in potentially dangerous situations, like the transport of relief supplies into crisis areas. Autonomous vehicles are reliant on systems recognising the environment around them to plan a route, but also to identify potential hazards. This work aims to help with the latter task by detecting hazardous substances in multispectral images of the vehicle’s surroundings as anomalies. Since training a supervised model would require an extensive labelled dataset, unsupervised methods will be explored instead. Unsupervised anomaly detection methods learn normal appearances by training on non-anomalous images and find anomalies as deviations from this normal state. A multispectral image dataset was recorded driving on roads for this purpose. In this work, two approaches are compared regarding their performance in the detection of potentially hazardous substance, a variational autoencoder (VAE) and a student-teacher based model. Evaluation of the model performance on the multispectral dataset revealed that the VAE’s performance is superior, while the student-teacher model performs better on additional RGB dataset, commonly encountered as benchmark in industrial anomaly detection. Experiments to determine if multispectral images offer benefits for this anomaly detection task remained inconclusive.