Vision-based 3D occupancy prediction in autonomous driving: a review and outlook (2405.02595v2)

Abstract: In recent years, autonomous driving has garnered escalating attention for its potential to relieve drivers' burdens and improve driving safety. Vision-based 3D occupancy prediction, which predicts the spatial occupancy status and semantics of 3D voxel grids around the autonomous vehicle from image inputs, is an emerging perception task suitable for cost-effective perception system of autonomous driving. Although numerous studies have demonstrated the greater advantages of 3D occupancy prediction over object-centric perception tasks, there is still a lack of a dedicated review focusing on this rapidly developing field. In this paper, we first introduce the background of vision-based 3D occupancy prediction and discuss the challenges in this task. Secondly, we conduct a comprehensive survey of the progress in vision-based 3D occupancy prediction from three aspects: feature enhancement, deployment friendliness and label efficiency, and provide an in-depth analysis of the potentials and challenges of each category of methods. Finally, we present a summary of prevailing research trends and propose some inspiring future outlooks. To provide a valuable reference for researchers, a regularly updated collection of related papers, datasets, and codes is organized at https://github.com/zya3d/Awesome-3D-Occupancy-Prediction.

• The paper provides a comprehensive review of vision-based 3D occupancy prediction methods using BEV, TPV, and voxel representations.
• The paper identifies key challenges in generating dense 3D occupancy annotations and evaluates metrics like mIoU and IoU with current datasets.
• The implications include promising future directions such as synthetic data generation, multi-agent collaboration, and integration of temporal dynamics.

Vision-based 3D Occupancy Prediction in Autonomous Driving: A Review and Outlook

Vision-based 3D occupancy prediction is emerging as a promising perception task in autonomous driving, providing a cost-effective approach to understanding the spatial occupancy and semantics of environments surrounding a vehicle. This paper presents a detailed examination of current approaches, challenges, and future directions for 3D occupancy prediction derived from image inputs.

Challenges in Vision-based 3D Occupancy Prediction

Task Definition and Ground Truth Generation

The primary task of vision-based 3D occupancy prediction is to classify each voxel in a 3D space based on camera inputs as either occupied or unoccupied, with additional semantic classification if occupied. Ground truth for this task is typically derived from LiDAR point clouds, but these are sparse and introduce challenges in generating dense annotations necessary for effective model training.

Figure 1: Visual comparison on 3D occupancy annotations.

Generating dense occupancy annotations typically involves fusing multi-frame LiDAR data and addressing the challenges posed by static and dynamic components of the scene. This complex process introduces additional computational and annotation challenges.

Datasets and Evaluation Metrics

Common datasets like SemanticKITTI and nuScenes provide foundational data, but 3D occupancy prediction demands fine-grained semantic detail not fully captured by these datasets. Current evaluation metrics such as Mean Intersection over Union (mIoU) and Intersection over Union (IoU) fall short of reflecting the detailed occupancy dynamics required for robust deployment of autonomous systems.

Taxonomy of Methods

The paper categorizes methods into feature enhancement, deployment-friendly, and label-efficient, each addressing specific challenges in 3D occupancy prediction.

Feature Enhancement Methods

These methods aim to improve the model's ability to discern 3D features from 2D inputs using different representations such as BEV, TPV, and direct operations on voxel representations.

Figure 2: Hierarchically-structured taxonomy of vision-based 3D occupancy prediction for autonomous driving.

• BEV-based Methods: Utilize bird's-eye view representations to extract spatial information, providing robustness against occlusion and depth ambiguities.

  Figure 3: Illustration of BEV-based methods.

• TPV-based Methods: Introduce tri-perspective views to enhance spatial understanding, allowing for a more comprehensive capture of 3D scene geometry.

  Figure 4: Illustration of TPV-based methods.

• Voxel-based Methods: Directly operate on 3D voxel grids for detailed feature extraction, capturing fine-grained spatial details.

  Figure 5: Illustration of Voxel-based methods.

Deployment-friendly Methods

These approaches prioritize computational efficiency, employing strategies such as perspective decomposition and coarse-to-fine refinement to reduce resource consumption while maintaining model fidelity.

Figure 6: FB-OCC \cite{li2023fb2} applies forward and backword projection to generate dense BEV features, which are unsqueezed for 3D occupancy prediction.

By exploiting view transformations and focusing computational efforts on critical areas or using less resource-intensive computation, these methods strive for real-time applicability.

Label-efficient Methods

Address the expense of annotating data for these tasks. Leveraging neural rendering techniques and unsupervised learning paradigms, these methods aim to eliminate or reduce dependency on labeled datasets.

Figure 7: Illustration of TPV-based methods.

Semantic guidance from rendered 2D views allows models to learn effectively without dense 3D supervision.

Future Outlook

Data Generation and World Models

Generating synthetic data using 3D occupancy frameworks offers a promising direction for augmenting training datasets without incurring high costs. Utilizing 3D occupancy in world models can enhance long-term prediction capabilities and dynamic scene understanding.

Multi-agent Collaboration

Collaborative perception across multiple vehicles could overcome limitations of single-agent systems in occlusion and range. Effective multi-agent frameworks can enable comprehensive environmental understanding by sharing perceptions across connected systems.

Task Integration

Future research should focus on integrating open-set recognition and 4D temporal dynamics into 3D occupancy frameworks. Combining spatial and temporal aspects with open vocabulary recognition will address the challenges of dynamic and evolving driving environments.

Conclusion

Vision-based 3D occupancy prediction continues to evolve, with promising strides in feature extraction, computational efficiency, and label efficiency. Addressing these challenges collectively and exploring synergies between methods can significantly aid in the advancement of autonomous driving technologies.