Machine Learning Methods for Analysis of Photo/Video Files from Cameras

Abstract

The research relevance is determined by the need to improve the accuracy and speed of image segmentation in complex urban environments for automated monitoring systems. The study aimed to develop and evaluate the effectiveness of deep neural networks for solving the problem of semantic segmentation of photo and video data obtained from surveillance cameras. The study tested DeepLabv3+ and U-Net architectures adapted for real-time image processing. Data augmentation methods, including adaptive contrast enhancement and illumination normalisation, have been implemented to improve the algorithms' resistance to adverse lighting conditions and weather factors. Experimental results on the Cityscapes and ADE20K datasets showed that the DeepLabv3+ model achieved an average IoU of 0.73 on the test data, and the use of optimised post-processing mechanisms reduced the accuracy drop to IoU = 0.65 in difficult shooting conditions. The image processing speed was 40 ms per frame, making the model suitable for use in real-time systems. The obtained results confirm the effectiveness of the proposed architectures and emphasise the need for further optimisation of the algorithms to improve the segmentation accuracy in low-light conditions. The practical significance of the study is to increase the reliability of traffic monitoring and public safety in the urban environment.