Current mainstream high-performance panoptic segmentation methods incur high computational overhead and are unable to meet the requirements of real-time applications.An innovative

single-stage

lightweight panoptic segmentation network based on camera sensors is proposed.The network employs a shared feature extraction backbone and is equipped with three dedicated branches for object detection

semantic segmentation

and instance-level attention mask generation

respectively.The designed novel panoptic segmentation head generates instance-aware soft attention masks through spatial embedding learning

achieving a natural fusion of object category semantic masks and instance soft masks at the pixel level

and directly outputs panoptic segmentation results without additional post-processing.Experiments on the COCO and Cityscapes datasets show that this method achieves state-of-the-art performance in both accuracy and speed.In particular

it achieves a panoptic quality score of 59.7 on the 1024×2048 resolution images of the Cityscapes dataset

while maintaining a real-time inference speed of over 10fps.