A Multi-Scale Feature Fusion-Based Model for High-Accuracy and Efficient Vehicle Detection in Complex Scenarios

Vehicle detection is essential in smart cities and intelligent transportation systems. However, current models face challenges in balancing detection accuracy and model complexity, limiting their deployment on low-performance edge devices, especially in complex scenarios with multi-scale objects and environmental variations. To overcome these issues, YOLO-MER, a novel lightweight multi-scale feature fusion vehicle detection network is proposed. First, a multi-scale feature fusion convolutional architecture, MERes Block, is developed based on an enhanced ResBlock to reconstruct the YOLO backbone, enabling feature extraction and fusion across four scales. Additionally, a lightweight neck architecture is designed to integrate features across three dimensions, ensuring comprehensive utilization of multi-scale information. The proposed YOLO-MER is validated on two public datasets. On the Vehicle dataset, it achieved the highest detection accuracy with 79.05% mAP, the smallest model size with 6.94 MB, and the fewest parameters with 1.59M. On the Crack dataset, it achieved the best detection accuracy with 86.11% mAP.