Leveraging Swin Transformer for Robust Object Detection in Autonomous Vehicles

Safety in diverse conditions is something which is extremely important for Autonomous driving systems. In this paper we explore the use of the Swin Transformer [1] – a state-of-the-art vision transformer with a hierarchical, shifted window design – as the backbone for object detection in camera-based autonomous vehicles. In this paper we have finetuned a pre-trained Swin Transformer on a the BDD100K [2] dataset, which includes a variety of images for example varied lighting and weather conditions. We have compared its performance against popular convolutional neural network (CNN) based detectors like Faster R-CNN [3] and YOLO [4], as well as the transformer-based DETR model [5]. Our experiments demonstrate that swin transformer significantly improves the detection accuracy and reliability under adverse conditions (such as low light, rain, and fog) while maintaining competitive computational efficiency. It achieves a higher mean average precision and better resilience to environmental changes due to its shifted window attention mechanism that ensures the computation cost remains manageable for real-time deployment. These results shows the promise of transformer based architecture in enhancing object detection for autonomous driving. We can conclude from this paper that using Swin Transformer enables us to have robust perception systems in self-driving vehicles and it outperforms the existing conventional CNN based detectors in both accuracy and robustness, without having prohibitive runtime overhead.