Quantum capsules in accident detection: A simple feature extraction on capsule networks for high-detail images

Abstrak

Accident detection from CCTV imagery remains challenging because complex traffic scenes contain multiple interacting objects and fine spatial structures whose representation quality may degrade under conventional feature extraction pipelines. To address this challenge, we propose QC-CapsNet, a hybrid quantum–classical Capsule Network model designed to enhance feature representation and investigate the routing process in high-complexity images for accident detection tasks. This model introduces quantum capsules, a feature representation that integrates parameterized quantum circuits (PQC) into the capsule transformation. Three new PQC variants are proposed as the main contribution, along with the pre-trained backbones, to improve feature map quality. Based on experimental results across several datasets, QC-CapsNet shows consistent performance gains over the evaluated baseline models, with margins of 0.17–10.90%. This model also competes with state-of-the-art methods, achieving an accuracy improvement of 0.69–1.81%. Overall, QC-CapsNet demonstrates the feasibility and potential of integrating PQC into capsule network architectures for high-complexity accident detection problems, while offering a foundation for further investigation of quantum layer architectures and designs in future work.

Identifier

• DOI: https://doi.org/10.1016/j.rineng.2026.112334
• OpenAlex: https://openalex.org/W7172393197