This paper presents a novel approach to Vehicle-to-Everything (V2X) communication, shifting the focus from transmitting raw sensor data to sending high-level "meaning." By extracting task-specific semantic information at the roadside, the authors demonstrate a way to support safe intersection navigation with minimal latency and massive bandwidth savings.
TL;DR
The researchers propose a goal-oriented V2X framework that extracts "intent" and "perception" semantics from roadside sensors to guide vehicles through intersections. By transmitting compact 12-byte packets instead of megabyte-scale video frames, the system achieves real-time performance and high decision accuracy while drastically reducing communication overhead.
What Problem Does This Paper Address?
Intersections are high-risk zones where a vehicle's onboard sensors are often limited by occlusions and a restricted field of view. While cooperative perception—where roadside infrastructure shares its "view" with vehicles—can solve this, traditional methods face a "data glut" problem.
Transmitting raw video or high-dimensional features from a Roadside Unit (RSU) to multiple vehicles consumes immense bandwidth and introduces significant latency. Furthermore, existing task-oriented V2X studies often focus on intermediate sub-tasks (like identifying a license plate) rather than the ultimate goal: safe passage. This paper addresses the need for a system that provides immediate, decision-level assistance (e.g., "Should I stop or go?") without overloading the communication channel.
Methods
The authors developed a transmitter-side semantic extraction pipeline that operates on the RSU. The system architecture is illustrated in Figure 2:

The pipeline consists of several key stages:
- Object Detection & Tracking: Using YOLO11s and BoT-SORT to identify and follow traffic participants in the RSU's video feed.
- Traffic Task Formulation: The researchers defined a set of "traffic tasks" based on Japanese traffic regulations. These tasks combine vehicle maneuvers (straight, left, right), signal states, and spatial regions (Regions of Interest or ROIs).
- Semantic Extraction: Instead of just "seeing" a car, the system determines its relation to ROIs and the traffic signal.
- Intent Inference: A Finite-State Machine (FSM) generates a "Recommended Action" (e.g.,
GO_STRAIGHT_LEFT,STOP,PREPARE) for the vehicle. - Packet Generation: All this information is compressed into a tiny, structured semantic packet.
Datasets
The framework was evaluated using a self-constructed dataset generated in the CARLA simulator.
- Environment: A four-way intersection under clear weather conditions.
- Scope: 2,250 frames of 1280x720 video at 15 fps.
- Scenarios: 12 distinct "ego-only" traffic tasks (T1–T12), categorized by maneuver and signal state, as detailed in the extracted table:
Key Insights and Ideas
- Goal-Oriented vs. Data-Oriented: The "aha moment" here is that a vehicle doesn't need to see the pixels of the intersection; it needs to know the state of the intersection relative to its goal. By shifting the objective to "task completion," the communication burden vanishes.
- Roadside Vantage Point: The RSU's high mounting point provides a "god's eye view" that is inherently better for extracting semantics than a vehicle-level sensor, which might be blocked by a truck.
- Intent as a Service: By moving the decision-making logic (the FSM) to the RSU, the system reduces the computational burden on the vehicle's onboard units, which only need to decode a simple packet.
Key Results
The experimental results validate the efficiency of the semantic approach:
- Accuracy: The system achieved a 100% task completion rate. The recommended action accuracy was nearly perfect, with minor errors only occurring during brief transitions between regions (see the confusion matrix in Figure 3).
- Latency: The mean total processing latency was 60.057 ms. This is critical because it stays below the 66.7 ms frame interval of the 15 fps input, ensuring the information remains valid when it reaches the vehicle.

- Compression: This is the most striking result. As shown in Figure 5, a raw image frame is roughly 1.28 MB, whereas the semantic packet is only 12 Bytes. This represents a reduction of five orders of magnitude.

Limitations and Future Work
While promising, the current study is a "first step" with acknowledged limitations:
- Simplified Scenarios: The evaluation focused on "ego-only" tasks (one vehicle at a time). Real-world intersections involve complex interactions between multiple vehicles, pedestrians, and cyclists.
- Environmental Factors: The dataset only covered clear weather. Performance in rain, fog, or nighttime remains to be tested.
- End-to-End Validation: The paper focuses on the transmitter side. Future work will integrate actual communication channels (like Visible Light Communication or RF) and measure receiver-side decoding latency.
- Unequal Error Protection: The authors plan to investigate how to protect the most critical parts of the semantic packet (like the "STOP" command) more robustly than less critical data.
Summary updated