Introduction
Connected and autonomous vehicles (CAVs) utilize cooperative perception to extend their sensing range by sharing data with nearby vehicles. However, this often leads to a "data deluge." If every vehicle broadcasts all detected objects, the network becomes congested, and the driver’s display—whether for a human or an AI—becomes cluttered with irrelevant information, such as parked cars or distant pedestrians.
AICP (Augmented Informative Cooperative Perception) is a fast-filtering system designed to address this by prioritizing "informativeness." It filters out redundant or irrelevant sensor data at both the network and application layers, ensuring that only the most critical safety information is processed and displayed.
System Architecture
The AICP framework is a full-stack solution that manages data from initial sensing to final display. The architecture is divided into three main pillars that operate across the sender and receiver pipelines.
- Vehicular Data Unit (VDU): This custom, lightweight data structure is designed for rapid interpretation. It encapsulates metadata like object type, position, and "creation time," allowing the system to quickly discard outdated or irrelevant information.
- Contextual Multihop Routing (CMR): Operating at the network layer, CMR filters packets based on geographic relevance. Instead of simple broadcasting, it uses directional routing to forward data only to vehicles traveling in a similar direction or within a specific hop limit, reducing redundant transmissions.
- Weighted Fitness Sorting: At the application layer, the system uses a near real-time sorting algorithm to decide which objects to render on an Augmented Reality Head-up Display (ARHUD). Using Mahalanobis distance, it ranks objects based on distance, velocity, direction, and category, prioritizing immediate threats like a pedestrian stepping into the street over static background objects.
Experimental Evaluation
The researchers evaluated AICP using both real-world prototypes and large-scale simulations:
- Real-world Prototype: A proof-of-concept was built using a Linux-based sender (running YOLOv5) and an Android-based receiver. Testing in Helsinki city center resulted in a dataset of over 75,000 objects.
- Large-scale Simulation: To test network performance, a 20 km² area of London city center was simulated using the Veins framework (OMNeT++ and SUMO), incorporating real-world traffic data to model various congestion levels.
Key Results and Insights
AICP demonstrates that informativeness is a decaying value; a message about a fast-moving vehicle loses relevance much faster than one about a stationary hazard. By applying the "7±2" rule of human cognition, the system limits the number of displayed objects to prevent information overload.
- Efficiency: The entire pipeline adds only 12.6 milliseconds of latency, meeting real-time safety requirements.
- Network Performance: In peak-hour simulations, AICP reduced the number of received Basic Safety Messages (BSMs) by over 80% compared to standard hop-based filtering, significantly decreasing Channel Busy Time.
- Visual Clarity: Qualitative results show a much cleaner ARHUD interface, highlighting critical objects while discarding background clutter, as seen in the comparison above.
Discussion and Future Work
While AICP effectively reduces data overload, the researchers noted that objects may occasionally "flicker" as the filter re-evaluates informativeness frame-by-frame. Future work will focus on implementing "object persistence delays" to stabilize the display and optimizing the placement of AR content to further minimize cognitive load. Additionally, deep learning could be integrated to predict informativeness in more complex, multi-vehicle accident scenarios.
Summary updated