Detecting accidents is also a central aspect of intelligent transportation system since the increasing number of vehicles on the roads also increases the chances of road accidents. The most recent advances in deep learning and transfer learning have made it possible to analyze traffic surveillance image in an automated manner. At the same time, quantum-classical machine learning has emerged as a possible paradigm to implement quantum circuits in the standard neural networks. This paper outlines a series of hybrid quantum transfer-learning architecture of image-based accident detection, whereby the pre-trained convolutional neural network (MobileNet, ResNet50V2, or EfficientNetV2S) operates as a feature extractor, and the final dense classifier is replaced by a parameterised quantum circuit (PQC) layer. The three PQC designs that use different qubit entanglement architecture designs are considered to help clarify the effect of circuit topology on classification behavior in data-scarce conditions. A dataset of experimental experiments on the closed circuit television (CCTV) accident images was used, which contained 989 images and multiple backbone architectures and optimisers were used. The findings show that the hybrid achieves performance comparable to classical dense-layer baselines and it uses significantly fewer parameters of the training phase of the classification step. In repeated trials, PQC based classifier also exhibits more consistent training dynamics and less variability in performance. These results indicate that hybrid quantum -transfer learning can become a parameter-efficient substitute classifier in the transfer-learning pipelines, especially in data-limited scenarios. However, the experiments were limited to one data set and based on quantum simulation; thus, the findings cannot be perceived as a definitive representation of the parameter-efficient hybrid architectures but as an investigation of possible improvements in predictive ability.
Sequential hybrid quantum transfer learning for robust and accurate accident detection
Pau, Giovanni
2026-01-01
Abstract
Detecting accidents is also a central aspect of intelligent transportation system since the increasing number of vehicles on the roads also increases the chances of road accidents. The most recent advances in deep learning and transfer learning have made it possible to analyze traffic surveillance image in an automated manner. At the same time, quantum-classical machine learning has emerged as a possible paradigm to implement quantum circuits in the standard neural networks. This paper outlines a series of hybrid quantum transfer-learning architecture of image-based accident detection, whereby the pre-trained convolutional neural network (MobileNet, ResNet50V2, or EfficientNetV2S) operates as a feature extractor, and the final dense classifier is replaced by a parameterised quantum circuit (PQC) layer. The three PQC designs that use different qubit entanglement architecture designs are considered to help clarify the effect of circuit topology on classification behavior in data-scarce conditions. A dataset of experimental experiments on the closed circuit television (CCTV) accident images was used, which contained 989 images and multiple backbone architectures and optimisers were used. The findings show that the hybrid achieves performance comparable to classical dense-layer baselines and it uses significantly fewer parameters of the training phase of the classification step. In repeated trials, PQC based classifier also exhibits more consistent training dynamics and less variability in performance. These results indicate that hybrid quantum -transfer learning can become a parameter-efficient substitute classifier in the transfer-learning pipelines, especially in data-limited scenarios. However, the experiments were limited to one data set and based on quantum simulation; thus, the findings cannot be perceived as a definitive representation of the parameter-efficient hybrid architectures but as an investigation of possible improvements in predictive ability.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
