Image Recognition in Autonomous Driving Based on Improved Swin Transformer

Engineering notes

We applied the model to SODAIOM, an autonomous driving dataset released by Huawei, and finally attained a classification accuracy of 95.3% and a detection mAP of 91.9%, both achieving state-of-the-art.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Traffic image recognition is one of the most important phases in the field of autonomous driving, including the classification of real-time periods and the detection of pedestrians, vehicles, etc. on the road. In this paper, we proposed an end-to-end classifcation and object detection method based on Swin Transformer with improved cascade RoI heads. Our method focuses on the scale problem from language to vision field in traditional Transformer model and the mismatch problem of bounding box regression in previous object detection methods (e.g. Faster R-CNN). A modified Swin Transformer architecture with multiple RoI heads is adopted in the proposed model to perform classification and object detection, meanwhile improved optimization strategies are used. We applied the model to SODAIOM, an autonomous driving dataset released by Huawei, and finally attained a classification accuracy of 95.3% and a detection mAP of 91.9%, both achieving state-of-the-art.

Links and sources

• Official / arXiv page

Comments