Detection of vehicles on images obtained from unmanned aerial vehicles using instance segmentation

Authors:

Abstract
In recent years, developing and improving of methods for automatic processing of imagery pictures taken from air with unmanned aerial vehicles (UAV) for ground vehicles detection have received priority. One of the most perspective approach in this direction is application of neural networks. However, UAV imagery has some distinctive features compared to normal images namely high resolution, complex image background, unequal distribution by size and quantity of training data, image bending due to flight-performance changes caused by air environment conditions. These factors compromise and make much more complicated detection of fine-graded objects by standard models of neural networks and in some cases renders automated processing of such images impossible. For the purpose of automatic detection of vehicles in images taken from UAV the analysis of existing methods of automated images processing and models of neural networks as the basis for these methods was conducted. The results of analysis allowed to highlight a multistage pipeline of aerial image, that combines approaches to detection, instance and semantic segmentation. The model of segmentation cascade was improved through considering objects geometrical dimensions and their ratio as well as scale changing and survey conditions. Application of suggested model of segmentation cascade for automated vehicle detection on aerial images allows to increase the accuracy of localization and detection of such objects.

This publication has 28 references indexed in Scilit: