Researchers have examined the challenge of detecting and classifying dynamic road obstacles for autonomous driving systems and presented a deep learning-driven convolutional neural network approach for the task. The review article focuses on mobile obstacles such as pedestrians, vehicles, bicycles, buses, trucks, cars, motorbikes, and animals, which can pose serious safety risks in complex driving environments.
Obstacle detection is one of the core perception tasks in autonomous driving. While autonomous vehicle technologies have advanced quickly in recent years, their safety impact remains an active area of research and debate. A vehicle that cannot reliably identify changing road hazards cannot make safe planning and control decisions, even if other parts of the system are technically sophisticated. For this reason, perception systems that detect and classify obstacles are central to the broader goal of improving road safety.
The article emphasizes that dynamic obstacles are especially important because they can move unpredictably and interact with the vehicle?s path in real time. Unlike fixed objects, pedestrians, animals, and moving vehicles can suddenly enter risk zones, change direction, or create collision hazards. Both drivers and the obstacles themselves may face fatal consequences if these risks are not recognized early. This creates a need for detection systems that can alert drivers or autonomous systems in advance.
In reviewing this problem space, the authors focus on deep learning methods, particularly convolutional neural networks, or CNNs. CNNs have become widely used in computer vision because they can learn visual patterns from image data and support classification and detection tasks. For autonomous driving, however, performance depends heavily on whether the training data reflect real road complexity, including diverse obstacle types, changing conditions, and visually similar object classes.
To support this need, the paper describes the creation of a comprehensive dataset that integrates data from various sources. According to the article, the dataset includes a diverse range of mobile obstacles, including pedestrians, several vehicle categories, and animals. This dataset-oriented contribution is important because robust obstacle detection depends not only on model design, but also on whether the model is exposed to the kinds of objects it will need to recognize in realistic driving scenarios.
The authors also present an advanced CNN-based detection and classification model, referred to in the article as OD-CNN-18 Layers. The model is designed to identify critical road risk factors and categorize detected obstacles in dynamic environments. This dual focus on detection and classification matters because an autonomous system must know both that an object exists and what kind of object it is. A pedestrian, motorbike, animal, and truck may require different risk assessments and driving responses.
The reported results indicate strong performance for the proposed CNN architecture. The paper states that the approach achieved a classification accuracy of 99.5% and a detection precision of 97.1%. These results suggest that the architecture can identify and classify road obstacles effectively under the evaluated conditions, contributing to improved situational awareness for autonomous driving systems.
At the same time, the review framing is important. Dynamic obstacle detection remains a broad and evolving research area, and performance in controlled datasets does not automatically guarantee universal reliability on all roads. Real deployment may involve lighting changes, weather, occlusion, sensor noise, unusual road users, regional traffic patterns, and edge cases that are difficult to represent fully in any dataset. Continued evaluation across larger and more varied scenarios will therefore be necessary.
Taken together, the article provides a useful contribution by connecting the safety motivation for dynamic obstacle detection with a CNN-based dataset and model framework. As autonomous driving technologies continue to develop, systems that can detect and classify moving road hazards accurately may play an important role in reducing collision risks and improving trust in intelligent transportation. The study suggests that carefully designed CNN models and diverse training datasets can help advance this goal, while also pointing to the need for continued validation in real-world conditions.
Reference
Author:
Hamza Assemlali, Soukaina Bouhsissin, Nawal Sael
Title of original paper:
Deep learning-driven CNN model for detection and classification of dynamic obstacles
Article link:
https://www.sciencedirect.com/science/article/pii/S2773153725000842
Journal:
Green Energy and Intelligent Transportation
DOI:
10.1016/j.geits.2025.100334
Affiliations:
Information Technology and Modeling Laboratory, Faculty of Sciences Ben M'Sik, Casablanca 20670, Morocco
Green Energy and Intelligent Transportation
Experimental study
Not applicable
Deep learning-driven CNN model for detection and classification of dynamic obstacles
3-Feb-2026