Abstract:Panoptic occupancy prediction using surround-view cameras is a core task in environmental understanding for autonomous driving. However, existing methods face challenges in simultaneously improving detection accuracy and maintaining computational efficiency. Traditional voxelization approaches introduce redundancy due to full-height dense computation, while feature compression may lead to the loss of fine-grained information along the height dimension. Moreover, multi-task coupling further degrades the prediction accuracy of small objects. To address this issue, this study proposes a lightweight network guided by dynamic sparse voxels, based on Panoptic-FlashOcc: (1) A dynamic sparse voxel sampling mechanism is designed to predict adaptive sampling points in the height dimension using a learnable mask, effectively reducing redundant computation. (2) A height-aware compensation module is introduced to recover spatial details via LSTM encoding and residual fusion. (3) A multi-task decoupled pyramid is constructed, employing deformable convolution to separate semantic and instance feature streams. On the Occ3D-nuScenes dataset, the proposed method improves the RayIoU metric by 5.5% over the baseline, achieving a score of 41.2%. Experimental results demonstrate that this approach significantly enhances small object detection accuracy and improves the real-time performance of panoptic occupancy prediction.