perception, prediction
The dangerous object is often not the one in view. It is the one a parked van is hiding.
Stand behind a supermarket endcap and you cannot see the cart coming from the next aisle. You still slow down, because experience says aisles hide carts. Driving has the same geometry. A bus hides a motorcycle. A hedge hides a driveway. A truck hides a person stepping out from the far bumper.
Occlusion is a visibility problem, not a labeling problem. No sensor sees through a van. A lidar point cloud of the van is still a van. The work is inferring that a hidden volume might contain something that will enter the lane in the next second.
You still slow down, because experience says aisles hide carts.
That inference is prediction wearing a perception costume. The system has to represent “unknown space,” not only named objects. If every unseen region is treated as empty, the stack is optimistic. If every unseen region is treated as occupied, the car never leaves a parking lot.
Good behavior around occlusion looks like covering a brake before the hidden thing appears. From the outside that can look psychic or overly timid. Inside, it is a bet about what kind of street produces surprises. Construction, school blocks, and parked-car corridors are where that bet is most expensive.