Multi-scale Feature Fusion Object Detection Based on Swin Transformer

The precision of object detection is crucial for autonomous vehicles to accurately perceive and understand the surrounding environment. In order to consider the problem of low detection accuracy caused by large differences in target scales and uneven content distribution, the paper proposes a multi-scale object detection method based on Swin Transformer. By enhancing the details of the combined features with low-level information, this method improves the effective utilization of features, thereby reducing the loss of crucial information when adjusting the number of channels. The algorithm presented in this paper improves the mAP by 0.83% and NDS by 0.93% on nuScenes dataset, and improves box AP by 1.4% on COCO dataset compared to the previous algorithm. Comparison results with other methods also validates the effectiveness of the improved algorithm.