Investigation of computer vision techniques for indoor navigation systems

Abstract

The subject of this article is the development and implementation of computer vision methods that can be integrated into an indoor navigation system designed for individuals with visual impairments. The goal of the study is to enhance such a system with advanced object recognition capabilities in enclosed environments by combining modern technologies, including artificial intelligence, spatial analysis, voice control, and Bluetooth-based localization. To achieve this, a number of tasks were carried out. These included an analysis of the problem domain and justification of the study’s relevance, a comparison of existing solutions, and the development of a generalized model of the navigation system with a voice interface, enabling real-time search for locations and items. A specialized dataset was prepared, containing images of key obstacle classes typically encountered in indoor environments – such as shopping carts, barrier tape, forklifts, and people. A new two-stage object recognition method was proposed to detect these classes in complex scenes. Additionally, a comparative analysis of deep learning architectures for object detection was conducted, followed by experimental studies to assess training quality and system robustness. The research employed various image preprocessing methods – bilateral filtering, Gaussian blurring, enhancement of specific color channels, motion blur removal, and noise reduction using averaging filters – as well as neural network-based methods for data analysis and statistical evaluation approaches. The results demonstrate that the proposed method significantly improves object detection performance on real-world images, achieving an average intersection-over-union (IoU) of 68% and a confidence level of 69%, which is 79% and 89% higher, respectively, compared to baseline recognition results on noisy inputs. However, the findings also revealed the necessity of integrating additional sensors, such as LiDAR, to reliably detect low-contrast or reflective obstacles like glass storefronts, which are difficult to identify using computer vision alone. Conclusions. The study confirms that the proposed two-stage preprocessing, and recognition pipeline significantly enhances navigation system performance for users with visual impairments, while also highlighting the importance of combining vision-based methods with complementary sensing technologies to ensure safe and reliable operation in complex indoor environments.