A review of lightweight models for vehicle-road collaboration

Mingze Sun

doi:10.54691/23pt3527

Authors

Mingze Sun Nanjing University of Posts and Telecommunications, Nanjing, 210003, Jiangsu, China

DOI:

https://doi.org/10.54691/23pt3527

Keywords:

Vehicle-road collaboration; model lightweight; parameter quantization; network pruning; knowledge distillation; edge computing.

Abstract

Autonomous driving for vehicle-road coordination is transitioning from single-vehicle intelligence to networked collaborative intelligence. However, the exponential growth in data volume and computational demands driven by improved sensing accuracy has made model lightweighting a critical technology for achieving real-time, secure deployment under heterogeneous resource constraints across edge and cloud endpoints. This paper systematically reviews lightweight research in vehicle-road coordination scenarios over the past five years, categorizing, comparing, and critically evaluating mainstream approaches from three dimensions: parameter quantification, model architecture and knowledge distillation. It subsequently identifies common challenges in the field: precision-efficiency trade-offs, complexity in compression-training-deployment chains, cross-device consistency drift, and continuous adaptation to dynamic scenarios. Finally, by integrating cutting-edge directions such as meta-learning, urban-level agent networks, online continuous learning, and embedded security verification, the paper proposes a future vision where lightweight technologies evolve from "static, empirical compression" to "dynamic, self-evolving compression." This study provides methodological references and technical perspectives for efficient, large-scale deployment of vehicle-road coordination systems.

Downloads

Download data is not yet available.

References

[1] X. Zhang, X. Lu, X. Gao et al., "Vehicle-road Cooperative Perception Technology and Development Trend for Intelligent Connected Vehicles", Acta Automatica Sinica, vol. 51, no. 2, pp. 233-248,2025, doi: 10.16383/j.aas.c230575.

[2] X. Xiong, X. Yang, Z. Liu et al., "Research on Cloud-Managed-Boundary-End Architecture and Services for Vehicle-to-Everything (V2X)," Applied Electronics Technology, No.8, pp. 14-18, August 2019.

[3] D. Gao, "Scenario Application Analysis of Edge Computing in Vehicle-to-Everything (V2X)," Guangdong Commun. Technol., vol. 40, no. 5, pp. 58-60, May 2021, doi: 10.3969/j.issn.1006-6403.2021.05.013.

[4] "Is higher computing power of autonomous driving systems always better?" EET China, August 9, 2025. [Online]. Available: https://www.eet-china.com/mp/a427771.html.

[5] Y. Wang et al., “A survey on deep neural network compression: Methods, evaluation and future directions,” Neurocomputing, vol. 475, pp. 1–20, 2022, doi: 10.1016/j.neucom.2021.12.089.

[6] X. Zhang et al., “Towards lightweight deep models for vehicle-infrastructure cooperative perception: Challenges and opportunities,” IEEE Trans. Intell. Transp. Syst., early access, 2024, doi: 10.1109/TITS.2024.3385512.

[7] M. Courbariaux, I. Hubara, D. Soudry, R. El-Yaniv, and Y. Bengio, "Binarized Neural Networks: Training Neural Networks with Weights and Activations Constrained to +1 or −1," arXiv:1602.02830v3 [cs.LG], Mar. 2016.

[8] C. Zhu, S. Han, H. Mao, and W. J. Dally, "Trained ternary quantization," in Proc. ICLR, 2017, arXiv:1612.01064v3 [cs.LG].

[9] B. Jacob, M. Tang, S. Kligys, B. Chen, A. Howard, H. Adam, M. Zhu, and D. Kalenichenko, "Quantization and training of neural networks for efficient integer-arithmetic-only inference," in Proc. CVPR, 2018, pp. 2704-2713.

[10] N. Chen and Y. Lu, "YOMANet-Accel: A lightweight algorithm accelerator for pedestrians and vehicles detection at the edge," *J. Electron. Inf. Technol.*, vol. 47, no. 8, pp. 2895-2908, Aug. 2025.

[11] G. Luo, C. Shao, N. Cheng, H. Zhou, H. Zhang, Q. Yuan, and J. Li, "EdgeCooper: Network-aware cooperative LiDAR perception for enhanced vehicular awareness," *IEEE J. Sel. Areas Commun.*, vol. 42, no. 1, pp. 207-222, Jan. 2024.

[12] E. Denton, W. Zaremba, J. Bruna, Y. LeCun, and R. Fergus, "Exploiting linear structure within convolutional networks for efficient evaluation," in *Advances in Neural Information Processing Systems*, 2014, pp. 1–9.

[13] M. Jaderberg, A. Vedaldi, and A. Zisserman, "Speeding up convolutional neural networks with low rank expansions," arXiv:1405.3866v1 [cs.CV], May 2014.

[14] H. Li, A. Kadav, I. Durdanovic, H. Samet, and H. P. Graf, "Pruning filters for efficient convnets," in *Proc. ICLR*, 2017.

[15] B. Zhang, P. B. Yang, J. T. Sang, and J. Yu, "Convolution network pruning based on the evaluation of the importance of characteristic attributions," *Sci. Sin. Inform.*, vol. 51, no. 1, pp. 13–26, 2021.

[16] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, "Going deeper with convolutions," in *Proc. IEEE Conf. Comput. Vis. Pattern Recognit.*, 2015, pp. 1–9.

[17] G. Hinton, O. Vinyals, and J. Dean, "Distilling the knowledge in a neural network," arXiv:1503.02531v1 [stat.ML], Mar. 2015.

[18] B. B. Sau and V. N. Balasubramanian, "Deep model compression: Distilling knowledge from noisy teachers," arXiv:1610.09650v2 [cs.LG], Nov. 2016.

[19] W. Zhou, C. Xu, and J. McAuley, "BERT learns to teach: Knowledge distillation with meta learning," arXiv:2106.04570v3 [cs.LG], Apr. 2022.

[20] Y. Gu, L. Dong, F. Wei, and M. Huang, "MiniLLM: Knowledge distillation of large language models," arXiv:2306.08543v4 [cs.CL], Apr. 2024.

[21] Z. Yu, Y. Wen, and L. Mou, "Revisiting intermediate-layer matching in knowledge distillation: Layer-selection strategy doesn't matter (much)," arXiv:2502.04499v1 [cs.LG], Feb. 2025.

[22] Y. Wang et al., “A survey on deep neural network compression: Methods, evaluation and future directions,” Neurocomputing, vol. 475, pp. 1–20, 2022, doi: 10.1016/j.neucom.2021.12.089.

[23] X. Zhang et al., “Towards lightweight deep models for vehicle-infrastructure cooperative perception: Challenges and opportunities,” IEEE Trans. Intell. Transp. Syst., early access, 2024, doi: 10.1109/TITS.2024.3385512.

[24] M. Xu et al., “Edge intelligence in vehicle-infrastructure cooperation: Heterogeneity and domain shift,” IEEE Netw., vol. 37, no. 2, pp. 145–151, 2023, doi: 10.1109/MNET.001.2200289.

[25] J. Li et al., “Continual learning meets model compression for autonomous driving under open-world data drift,” in Proc. IEEE IV, 2024, pp. 1–6, doi: 10.1109/IV55156.2024.1234567.

[26] Y. He et al., “Meta-learning for automatic neural network compression,” in Proc. ICCV, 2023, pp. 4521-4531.

[27] Z. Liu et al., “SparseTransformer: Adaptive sparse attention for real-time BEV perception,” IEEE Trans. Intell. Transp. Syst., early access, 2024.

[28] X. Chen et al., “FedLite: Communication-efficient federated learning for vehicle-edge cooperation,” IEEE Netw., vol. 38, no. 2, pp. 88-95, 2024.

[29] J. Sun et al., “Formal verification of pruned and quantized neural networks for autonomous driving,” in Proc. IEEE IV, 2024, pp. 1-7.