Infrared-based object classification for the surveillance of valuable infrastructure

Abstract

The surveillance of valuable infrastructure, such as photovoltaic parks, is considered of fundamental importance for their proper function and maintenance as well as the avoidance of criminal damage incidents. At the same time, the privacy of employees working in the same area should not be jeopardized and their personal data should always be protected. The use of thermal cameras presents a solution to both of the above issues by offering an unobtrusive surveillance approach with the ability to supervise industrial premises under a wide range of environmental and situational conditions. The current paper proposes an algorithm for the classification of moving objects that aims to increase the efficiency of surveillance methodologies by shifting the focus on high-risk classes, such as humans instead of animals. The proposed methodology utilizes an automated decision framework that determines when textural features are fit to be used, based on the discriminative power of the texture of the object. Many texture descriptors were tested, including Local Phase Quantisation and Histograms of Oriented Gradients, resulting in the use of a lately proposed combination of these descriptors. This new multi-class object classification approach introduces the use of confidence values and a voting system to achieve a more accurate selection of the appropriate class. The velocity was also used as a discriminative feature, especially to help distinguish between humans and motorcycles. Several algorithms have been used to validate the results of the experimental studies with special focus on the classification accuracy. The experimental results were obtained from a series of scenarios demonstrated in four different condition sets (different temperature-humidity-illumination), that exposes the advantages and disadvantages of the proposed unimodal classification method in infrared imagery. The dataset is also benchmarked against another state-of-the-art approach.