Abstract
Detection of liquid spill is a crucial and effective task to maintain safety and protection in various environments. Thermal imaging as a passive imaging modality working in different lighting conditions and even through smoke can be advantageously used to detect liquid spill in challenging conditions. Deep learning-based object detectors are well-established techniques to detect and localize different objects or phenomena in a variety of image modalities, however they require large scale databases with bounding box annotation in order to be trained from scratch. In this work, we present, evaluate, and compare three different methods to address the unavailability of substantial datasets dedicated to liquid spill detection from thermal images in the context of health and safety prevention. A Flir A35 thermal camera is used to collect data for the experiments. The three methods are based respectively on a conventional image processing algorithm using watershed segmentation, a weakly supervised approach using Gradient Class Activation Mapping, and an unsupervised deep learning approach for salient object detection guided by motion. No pixel level annotation is required for the proposed approaches. The work demonstrates that a conventional image processing approach, achieving an average precision and an average recall as high as 0.83 and 0.72 respectively, can reliably detect and localize warm liquid spill in sequences of thermal images.