Abstract
In mixed traffic flow scenarios, multiple types of traffic participants coexist on the same roadway, posing severe challenges for object detection algorithms due to significant disparities in target scales, complex background interference, dense occlusions, and the high heterogeneity of classes. Existing CNN-based detectors are constrained by the fixed receptive fields inherent in convolution operations and are generally plagued by imbalances between positive and negative samples as well as inadequate representations of small objects, further limiting their performance in mixed traffic detection tasks. To address these issues, we propose the MTF-NET detection network, which is endowed with full-field perceptual capabilities. First, a combination of CNN and MetaFormer is employed as the backbone for feature extraction to enhance contextual modeling. Second, to mitigate the inherent dual-dimensional information loss and small-target representation bottlenecks associated with pyramid structures, we introduce a Hierarchical Implicit-Explicit Pyramid structure alongside a Multi-Kernel Dilation Fusion Network designed to counteract the information degradation brought about by pooling operations. Finally, the Dynamic Dual Detection Heads utilize a dual-branch design that facilitates end-to-end deployment while alleviating the limitations imposed by non-maximum suppression (NMS), and a hybrid strategy integrating Exponential Adaptive Loss with Focaler-DIoU is developed to address the imbalance between positive and negative samples across multiple classes. Experimental results demonstrate that MTF-NET achieves a 5.1% improvement in mAP50 on the VisDrone2019 dataset, surpassing current state-of-the-art methods, and further yields enhancements of 4.2% and 13.4% on the UA-DETRAC-G2 and HazyDet datasets, respectively. These findings effectively validate the robustness and generalization capabilities of our network, providing a potent solution for object detection in complex mixed traffic flow scenarios.