ISSUES OF IMPROVING THE EFFICIENCY OF EMBEDDED CONTROL SYSTEMS WITH A COMPUTER VISION MODULE
DOI:
https://doi.org/10.34185/1991-7848.itmm.2026.01.070Keywords:
computer vision, embedded systems, automatic control, TinyML, neural network compressionAbstract
This paper investigates the relevant problem of improving the efficiency of embedded automatic control systems, where deep learning-based computer vision algorithms act as virtual sensors. Methods of model compression (quantization, pruning) within Edge AI and TinyML approaches are analyzed. It is hypothesized that isolated software optimization of neural networks, without considering the dynamics of the physical object, may reduce the system's stability margins due to the emergence of stochastic noise and variable latency. A comprehensive research methodology is proposed, combining the use of digital twins, simulation (Vision-in-the-loop), emulation, and hardware implementation (Hardware-in-the-loop). This approach aims to find the optimal trade-off between inference latency and recognition accuracy to enhance the reliability and efficiency of autonomous cyber-physical systems.
References
El Zeinaty, C., Hamidouche, W., Herrou, G., & Menard, D. (2024). Designing object detection models for TinyML: Foundations, comparative analysis, challenges, and emerging solutions. ACM Computing Surveys, 56, 1–46. https://doi.org/10.1145/3744339
Park, J., Kim, P., & Ko, D. (2025). Real-time open-vocabulary perception for mobile robots on edge devices: A systematic analysis of the accuracy-latency trade-off. Frontiers in Robotics and AI, 12, 1693988. https://doi.org/10.3389/frobt.2025.1693988
Khudiakov, I. V., Gritsuk, I. V., Chernenko, V. V., et al. (2021). Osoblyvosti modeliuvannia ta pobudovy informatsiinoi systemy dystantsiinoho monitorynhu tekhnichnoho stanu transportnykh zasobiv. Visnyk mashynobuduvannia ta transportu, 14(2), 140–148. https://doi.org/10.31649/2413-4503-2021-14-2-140-148.
