Dual-aggregation feature compilation network for urban traffic object detection and pedestrian pose estimation

With the increasing complexity of urban transportation systems, object detection and pedestrian pose estimation play a crucial role in intelligent traffic management and autonomous driving technologies. However, existing feature compilation networks are often designed for single tasks and perform poorly in small object detection and high occlusion pedestrian pose estimation tasks. To address the above issues, this paper proposes an efficient feature compilation network with Dual-aggregation, compatible with both object detection and pedestrian pose estimation. This network adopts a transfer learning-like training strategy in the feature extraction network, using a micro-complex convolution structure during training to bring the training results as close as possible to global optimization. During inference, a single simple convolution is used to inherit the training results, improving the model performance while ensuring model lightweight. The feature fusion employs a global-local dual aggregation structure, simultaneously considering multi-scale global and local features. Additionally, we use multiple public datasets to create a hybrid dataset under various scenarios to validate the robustness of the network. The experiments show that the proposed method outperforms existing mainstream methods in detection accuracy for urban object detection and pedestrian pose estimation tasks, especially demonstrating better robustness in complex urban traffic scenarios.