TRANSFER LEARNING SOFTWARE SYSTEM FOR ROAD RISK ASSESSMENT BASED ON OBJECT AND HUMAN BEHAVIOURAL FEATURES

Анотація

This paper introduces a software system designed for transfer learning, specifically for evaluating road risks based on object and human behavioural features under the limited computational resources of edge devices. The relevance of this study stems from the fact that modern road scene analysis systems are primarily focused on object detection and scene understanding, yet they insufficiently account for the behavioural characteristics of traffic participants and the specifics of transferring behavioural risk representations into compact models for online deployment.

The aim of this work is to develop a transfer learning software system for real-time online assessment of road risks based on the integration of object-based and behavioural features. A road scene representation model is proposed, which includes spatial, trajectory, and behavioural characteristics, as well as a mathematical model for risk assessment within the Safe–Suspect–Unsafe classes.

A teacher–student transfer learning model with a behaviour-aware loss function is developed that, unlike traditional knowledge distillation approaches, preserves risk-oriented behavioural knowledge. A software architecture integrating object detection, feature extraction, behavioural analysis, transfer learning and risk inference for online road risk monitoring is proposed.

Experimental results have shown that behavioural-aware transfer learning improves the accuracy of risk assessment, enhances the detection of critical scenarios, and ensures high consistency of predictions after knowledge distillation while maintaining suitability for deployment on resource-constrained platforms.

The proposed contribution consists in the combination of behavioural features, transfer learning, and risk-oriented scene reasoning within a unified software system for road risk assessment. The practical value of the results lies in the possibility of using the proposed approach in intelligent transportation systems and systems for monitoring hazardous road situations.

References

1.           Awan, A. A., et al. (2025). Deep learning methods for scene understanding in autonomous driving: A review. Expert Systems with Applications, 254, 128098. DOI: 10.1016/j.eswa.2025.128098

2.           Jiao, X., et al. (2024). RS-YOLO for efficient road scene object detection. Digital Signal Processing, 154, 104889. DOI: 10.1016/j.dsp.2024.104889

3.           Niu, Y., et al. (2024). Important object detection through driving scene relationship understanding. IEEE Access, 12. DOI: 10.1109/ACCESS.2024.3400589

4.           Tian, Y., et al. (2023). Semantic relationships in urban traffic scenes. IEEE Open Journal of Intelligent Transportation Systems, 4. DOI: 10.1109/OJITS.2023.3260624

5.           Fang, Z., et al. (2024). Behavioral intention prediction in driving scenarios: A survey. IEEE Transactions on Intelligent Transportation Systems. DOI: 10.1109/TITS.2024.3374342

6.           Xie, H., et al. (2025). Human fuzzy semantic logic for driving risk analysis. Accident Analysis & Prevention, 197, 108037. DOI: 10.1016/j.aap.2025.108037

7.           Guo, X., Zhou, W., & Liu, T. (2024). Contrastive learning-based knowledge distillation for urban scene segmentation. Knowledge-Based Systems, 292, 111588. DOI: 10.1016/j.knosys.2024.111588

8.           Gao, Y., et al. (2025). Domain adaptation for semantic segmentation of road scenes. Neurocomputing, 613. DOI: 10.1016/j.neucom.2024.128744

9.           Rakytyanska, H. (2023). Knowledge distillation in granular fuzzy models. In LNDECT. Springer. DOI: 10.1007/978-3-031-32095-8_4

10.        Rakytyanska, H., & Prus, B. (2024). Constructing prototype-based granular fuzzy rules for scene classification on mobile devices. In LNDECT (Vol. 219, pp. 194–218). DOI: 10.1007/978-3-031-70959-3_10

11.        Yao, Y. (2018). Three-way decision and granular computing. International Journal of Approximate Reasoning, 103, 107–123. DOI: 10.1016/j.ijar.2018.09.005

12.        Pedrycz, W. (2018). Granular Computing. CRC Press. DOI: 10.1201/9781315216737

13.        Angelov, P., & Gu, X. (2018). Deep rule-based classifier with human-level performance. Information Sciences, 463–464, 196–213. DOI: 10.1016/j.ins.2018.06.048

14.        Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network. arXiv:1503.02531. DOI: 10.48550/arXiv.1503.02531

15.        Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time object detection. In CVPR (pp. 779–788). DOI: 10.1109/CVPR.2016.91

16.        Gou, J., et al. (2021). Knowledge distillation: A survey. International Journal of Computer Vision, 129(6), 1789–1819. DOI: 10.1007/s11263-021-01453-z

17.        Mazhar, S., et al. (2023). Lightweight semantic segmentation for resource-constrained devices. Engineering Applications of Artificial Intelligence, 126, 107086. DOI: 10.1016/j.engappai.2023.107086

18.        Ma, S., et al. (2024). Weakly supervised object localization via knowledge distillation. Neurocomputing, 576, 127213. DOI: 10.1016/j.neucom.2023.127213

19.        Cordts, M., et al. (2016). The Cityscapes dataset for semantic urban scene understanding. In CVPR (pp. 3213–3223). DOI: 10.1109/CVPR.2016.350

20.        Yu, F., et al. (2020). BDD100K: A diverse driving dataset. In CVPR (pp. 2636–2645). DOI: 10.1109/CVPR42600.2020.00271