METHODOLOGY FOR DETECTION OF SMALL-SIZED COMPACT OBJECTS BASED ON COMPLEXING AERIAL SURVEYING DATA OF VARIOUS PHYSICAL NATURE

Abstract

The article is devoted to the development of a method for detecting small-sized compact objects, namely anti-tank and anti-personnel surface mines, based on the integration of aerial survey data of various physical nature. The relevance of the study is due to the acuteness of the problem of mine contamination of the territory of Ukraine, which is catastrophically increasing due to the development of unmanned and robotic mining vehicles. Although methods for detecting mines based on aerial survey data are intensively developing, complete descriptions of the method are rarely published so far, and mostly they concern individual technologies. The proposed method is based on the integration of detection results based on data of various physical nature with a complex differentiated approach to the use of machine and deep learning methods for data types of various physical nature, namely - visible range images, thermal images and magnetometry data. The priority of the selected data types was determined, and for visible range images, an approach was developed to minimize the required amount of training data and training time to ensure the ability to overcome the shortage of training data and quickly iteratively improve detection capabilities, including when expanding the set of search object classes. To achieve this goal, transfer learning (fine tuning) was used on the basis of large sets of public data with an emphasis on achieving optimal transfer learning hyperparameters for the base and target data sets and a specific neural network model, namely the number of training epochs and the depth of freezing of the layers of neural network coefficients. For thermal images, the possibility of their use in deep learning was proven both independently and in combination with visible range results. For magnetometry data, the feasibility of statistical methods for detecting anomalies and using the results to improve the reliability of detection through combination with detection results in images was demonstrated. Data obtained on test sites using real and training mine samples was used to test the developed methodology.