Deep Learning-Based Multi-Class Vehicle Detection: A High-Speed Approach for Traffic Surveillance and Smart City Applications
The rapid growth of urbanization requires effective traffic surveillance systems designed for smart city use. This paper presents a new Multi-Class Vehicle Detection Framework (MCVD-T) that uses Vision Transformers (ViTs) and improved feature extraction techniques for real-time, precise multi-class vehicle detection. Unlike previous models, such as CNNs and YOLOv9, this framework utilizes Swin Transformers, which are able to perform feature extraction at multiple levels. The implementation of Swin Transformers makes it feasible to operate in dynamic and complex environments. To lessen the computational load, which can significantly hinder performance in real-time applications, the framework applies lightweight attention mechanisms and model pruning techniques to provide better resource utilization, thus lessening inhibit delay and improving scalability in large-scale traffic networks. Additionally, knowledge distillation helps lessen the computing effort while providing a smaller yet effective inference model. The experimental results confirm that the MCVD-T framework surpasses the best existing methods by 22% improved detection accuracy and a processing time reduction of 35%. Both offline and online traffic monitoring capabilities are provided by this system, which allows for real-time insights using a fast data pipeline embedded within edge sensors. Its online functionality is asynchronous streaming based, which guarantees that video is analyzed consistently and allows for rapid vehicle classification for traffic management and law enforcement. MCVD-T architecture overcomes the common problems associated with such a system and provides a fast, scalable, and configurable solution, making it a critical part of smart urban infrastructure. It integrates seamlessly into smart city applications and enhances traffic management, public safety, and more mobile options for efficiency.