Recently, a breakthrough has been made in the field of intelligent driving perception by the team led by Dr. Jiamao Li, a researcher at the Laboratory of Biomimetic Vision Systems at the Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, in collaboration with partners. Two achievements in grid occupancy prediction and panorama segmentation have been accepted by ICRA.
Addressing the issue of grid occupancy prediction, the research team proposed a novel approach named CVFormer, based on the baseline algorithm (TPVFormer). This method adopts a surround-view perspective centered on the ego vehicle to represent occupancy prediction. By utilizing panoramic surround-view multi-perspective input, it characterizes the three-dimensional scene from orthogonal viewpoints around the vehicle, effectively achieving fine-grained 3D scene representation and overcoming the problem of viewpoint occlusion by obstacles around the vehicle body. The team validated this method on the authoritative driving scene occupancy prediction benchmark dataset nuScenes, demonstrating significantly improved prediction accuracy over existing algorithms while significantly reducing computational complexity, thereby exhibiting strong deployability. Additionally, this method has undergone testing and verification on intelligent driving perception systems.
The grid occupancy prediction CVFormer achieves optimal performance on the nuScenes dataset.
For the panoramic segmentation task, the team has devised an end-to-end panoramic segmentation model called BEE-Net, based on gated encoding and edge-constrained mechanisms. This model effectively addresses the semantic-instance prediction conflicts and edge segmentation challenges in panoramic segmentation through global bidirectional information interaction and multi-angle edge optimization. The approach has been validated on the authoritative CityScapes dataset for driving scene segmentation, surpassing existing CNN-based panoramic segmentation models in accuracy while being more efficient than all Transformer-based panoramic segmentation models. Additionally, the method balances the performance requirements of segmentation accuracy and efficiency and has been tested and verified in intelligent driving perception systems.
BEE-Net achieves state-of-the-art performance for CNN-based semantic segmentation on the CityScapes dataset. All images are sourced from Microsystems.
