The Occupancy Perception Network (also known as the Occupancy Network or OCC for short) is a pivotal technology in today's autonomous driving landscape. Given the dynamic nature of autonomous driving scenarios, most objects perceived are in constant motion. Therefore, any discussion of the occupancy perception network must be rooted in its ability to detect and interpret dynamic objects.

When delving into this topic, it's essential to broaden our view from simple 3D reconstruction to 4D spatiotemporal perception. The challenge with dynamic objects stems from their ever-changing positions, shapes, and speeds over time. OCC does not merely rely on frame-by-frame comparisons to track these changes; instead, it employs a sophisticated mechanism of temporal feature fusion and motion vector prediction to deeply model the dynamic attributes of the physical world.

How Are Temporal Features Aligned Across Grid Points?

The initial step in processing dynamic objects involves establishing a unified time reference frame. As an autonomous vehicle navigates, its coordinate system continuously shifts, meaning that the same grid point observed at different times will occupy inconsistent positions in image or feature space. To enable the occupancy perception network to comprehend object motion, vehicle motion compensation must first be implemented.

The system leverages the vehicle's inertial navigation data and odometry information to project feature maps from historical moments into the coordinate system of the current frame. This process employs feature alignment techniques, which translate and rotate features from multiple past frames in 3D space to align stationary backgrounds across the temporal dimension. Once the background is aligned, feature points that have moved in space become clearly visible, providing the network with a foundation for perceiving changes.

During the feature fusion stage, the OCC architecture utilizes 3D convolutions or temporal attention mechanisms. The network not only extracts current geometric features but also reviews feature sequences from hundreds of milliseconds or even longer in the past. This multi-frame fusion approach enables the network to transcend the limitations of single-frame images and capture the continuity of object motion. Even if an object appears blurred in a particular frame due to lighting or occlusion, accumulated features from historical frames provide effective supplementation, ensuring smooth and stable perception results.

How Does Occupancy Flow Quantify Object Motion?

Merely knowing that a grid point is moving is insufficient; the system must accurately determine its speed, direction, and magnitude. Within the OCC technical framework, this is achieved by outputting occupancy flow. Each small grid block marked as occupied not only stores the probability of an obstacle at that location but also carries a three-dimensional motion vector.

The generation of this motion vector relies on a dedicated prediction branch. At the backend of the network, the algorithm calculates the correlation between the current grid point and its historical counterparts to derive the instantaneous displacement of the grid point in 3D space. This means that for every vehicle and pedestrian on the road, OCC does not output a single overall motion value but rather motion vectors for the thousands of tiny grid points that compose these objects. This grid-level speed representation can describe highly nuanced object dynamics, such as speed variations across different parts of a vehicle during a turn or local movements of a pedestrian's arms.

This approach bypasses the complex target tracking stage found in traditional perception methods. In traditional schemes, if tracking is lost, speed information is also lost; whereas in OCC, as long as space remains occupied, speed vectors can be continuously output through temporal features. This logic, which directly maps low-level pixel features to physical motion attributes, significantly enhances the system's adaptability to irregular objects and complex motions because the network no longer attempts to understand who is moving but instead calculates how space is changing.

How Is Dynamic Prediction Maintained Under Occlusion?

The most challenging scenario in dynamic object perception is when objects disappear from view or are partially occluded. The core of OCC's ability to handle such issues lies in its spatiotemporal consistency modeling capability. When a dynamic object enters an occluded area, current sensor data cannot provide its positional information, but the network's internal temporal encoder retains the object's state features.

By incorporating spatiotemporal attention mechanisms, the network can learn the inertial patterns of physical motion. When processing temporal feature sequences, the attention mechanism assigns weights to feature points with strong motion trends. Even if the current frame's input is empty, the network can still generate predictions at potential occupied locations based on the occupied states and velocity vectors from previous frames. This is akin to equipping the perception system with a predictive brain, enabling it to infer the object's spatial distribution over the next second or two based on its trajectory before disappearance.

This prediction is not a blind guess but is based on probabilistic reasoning. The system outputs an occupancy probability map that gradually diffuses over time, indicating regions where the object may appear. This approach significantly optimizes the safety of autonomous driving, as the planning and control system can proactively avoid these highly probable occupied spaces without waiting for the object to fully reappear in view. This deep exploration of spatiotemporal continuity is precisely what gives OCC greater safety potential compared to traditional detection schemes.

What Changes Does Full-Scenario Dynamic Perception Bring?

This grid-based dynamic processing scheme has fundamentally transformed the efficiency of autonomous driving in handling complex road conditions. In traditional workflows, perception, tracking, and prediction are three independent stages, with errors from each stage accumulating. OCC integrates these functions within an end-to-end framework, directly outputting a three-dimensional spatial map with motion attributes. This highly integrated approach not only reduces computational latency but also eliminates perception interruptions caused by target matching errors.

For downstream decision-making and planning, this perception result is highly advantageous. The planning and control algorithm no longer needs to process lists of hundreds or thousands of targets but instead faces a real-time updated dynamic three-dimensional grid map with velocity information. This map clearly indicates which spaces are absolutely safe and which spaces will be occupied by dynamic objects in the near future.

This advancement in perception logic enables autonomous driving systems to handle unexpected situations more calmly. Whether it's a food delivery vehicle suddenly darting out from the roadside or scattered and sliding cargo ahead, OCC can capture and process these events using a unified logic. This modeling approach, which captures the most primitive and essential aspects of the physical world, is becoming a crucial technical support for achieving high-level autonomous driving capabilities, allowing vehicles to gain a more precise and stable sense of spatial control in the ever-changing urban traffic environment.

-- END --