Occupancy Network

An Occupancy Network is a deep learning model that elevates autonomous driving perception from merely "identifying objects" to "perceiving spatial geometry". Instead of trying to determine exactly what an obstacle ahead is, it divides the surrounding 3D space into countless micro cubes (known as voxels), subsequently assessing voxel by voxel whether each unit is "occupied" by an object or "free".

🧠 Core Concept: From "What is it?" to "What is there?"

Conventional perception systems rely heavily on object detection, requiring the identification of specific object categories (such as cars or pedestrians) before planning evasive manoeuvres. However, when encountering irregularly shaped obstacles absent from the training dataset—such as fallen cardboard boxes or overturned lorries—the system can easily suffer from blind spots.

Occupancy Networks completely overhaul this approach. They do not concern themselves with what an object is, but rather the spatial state itself. As long as an entity occupies physical space—regardless of how unusual its shape is or whether the system has seen it before—it is flagged as an obstacle, fundamentally eliminating missed detections in "long-tail scenarios" that plague traditional perception systems.

⚙️ Key Implementation Principles

• Spatial Voxelisation: The 3D space surrounding the vehicle within a set perimeter is partitioned into a grid of voxels, typically measuring 0.2 to 0.5 metres per side, creating a "3D pixel" map.

• Sensor Input and Fusion: Mainstream setups generally use feeds from surround-view camera systems as a foundation, inferring the occupancy probability of each voxel via neural networks. However, this typically requires high-precision data generated by LiDAR during early-stage training to supervise the vision model. In real-world applications, carmakers like Huawei and Li Auto also deploy LiDAR for real-time perception, fusing it deeply with vision to boost reliability in challenging scenarios like pitch darkness or strong glare.

• High Computational Demands and Challenges: Processing massive streams of 3D voxel data in real time places immense strain on in-vehicle chip computing power. Consequently, large-scale commercial deployment hinges on potent compute platforms alongside algorithmic optimisations—such as non-uniform voxel grids—to strike a balance between precision and processing efficiency.

🚀 Mainstream Technological Routes

In summary, by digitising the physical environment into a computable 3D spatial model, Occupancy Networks endow autonomous vehicles with genuine "spatial intuition". It has established itself as an indispensable puzzle piece in the journey towards high-level autonomous driving.