← Back to selected research
Research Case Study · Connected & Automated Mobility

Scenario-Adaptive Multi-Vehicle Cooperative Perception

Our team compared cooperative perception models across urban and highway scenarios. Using CARLA, SUMO, and OpenCDA, we examined how traffic conditions and sensing configurations affect reliability, then explored model selection and parameter tuning.

InstitutionTongji University
PeriodNov 2022 – Apr 2024
EnvironmentCARLA · SUMO · OpenCDA
FocusBEV cooperative perception · NAS
30Simulation configurations
13BEV perception models evaluated
5Urban & highway scenarios
Original cooperative perception result from the project presentation
Original project artifact · cooperative perception output
Team project at Tongji University · I extracted evaluation data from raw simulation files and compared reliability across the tested sensor configurations and perception schemes.

How does cooperative perception perform when the road environment changes?

A vehicle can miss objects because of occlusion or limited sensing range. Sharing information with nearby vehicles and roadside sensors can extend its view. A bird’s-eye-view (BEV) representation puts these observations into a common spatial frame.

We studied how accuracy and reliability change across road environments, and whether model choice, sensing configuration, and training settings should change with them.

Original project figure
Vehicle camera configuration
Figure 1. Example vehicle-mounted camera arrangement used as a reference for multi-view perception.
Original project figure
Roadside sensor layout
Figure 2. Example roadside sensing layout. The project compared lower- and higher-configuration infrastructure sensing schemes.

Comparing 30 simulation configurations.

The simulation set covered both urban and highway operation:

  • Highway basic segment
  • Highway curve
  • Highway ramp
  • Urban basic road segment
  • Urban intersection

Highway scenarios used a 100 km/h design reference, while urban segments and intersections used lower design speeds. Three traffic-flow levels were introduced to represent free-flow, stable-flow, and saturated conditions.

Sensor setup

Each vehicle used six RGB cameras (70° field of view, 1920 × 1080): front, front-left, front-right, rear, rear-left, and rear-right. Roadside sensing was tested under low- and high-configuration schemes.

Original project figure
Six camera views used in the cooperative perception simulation
Figure 3. Representative six-direction vehicle-camera views used in the simulated perception environment.

Combining traffic simulation and perception models.

CARLAConstruct road scenes, vehicle environments, sensor viewpoints, and visual perception inputs.
SUMOControl traffic demand, design speed, and traffic-flow conditions for each scenario.
OpenCDAIntegrate cooperative-driving modules and synchronize perception and motion-state data.
BEV perceptionGenerate cooperative bird’s-eye-view perception results from multiple sensing agents.
Reliability analysisEvaluate loss, uncertainty, and scenario-dependent performance before optimization.
Original project figure
Cooperative BEV perception pipeline
Figure 4. A BEV cooperative perception pipeline used as a methodological reference in the project.

Evaluating uncertainty alongside perception accuracy.

The project constructed an uncertainty-prediction framework that separates two forms of uncertainty:

Aleatoric uncertainty

Uncertainty associated with the data and sensing task, including noise and scene ambiguity.

Epistemic uncertainty

Uncertainty associated with model parameters and limited knowledge of unfamiliar conditions.

Reliability was then examined from both algorithm and vehicle perspectives, with vehicle-level failures analyzed across temporal and spatial dimensions. The resulting loss / reliability measure became the basis for comparing candidate perception models across scenarios.

Comparing 13 models before tuning a selected baseline.

For each scenario, the collected data were used to train and compare 13 BEV-oriented cooperative perception models. The model with the smallest perception loss was treated as the scenario-specific baseline for further optimization.

Neural Architecture Search (NAS) was then used to explore hyperparameter settings such as learning rate, dropout, and training duration.

  • Adaptation to known scenarios: select a model and parameter set that performs well for an identified traffic environment.
  • Evolution under unfamiliar scenarios: search for improved settings when the existing model is no longer sufficiently reliable.
Hardware / sensing adaptation

The study also compared roadside-sensor configurations by scenario, linking perception performance to sensing design rather than treating model architecture as the only optimization variable.

Optimization reduced the reported perception loss by approximately 99.98% in the selected comparison.

Before optimizationAfter optimization
Modelciassd_intermediate_fusionpoint_pillar_cobevt
Training epochs15170
Perception loss29.29365560.0056841

The project report also records improved average perception accuracy across multiple IoU thresholds after optimization.

Original project figure
Cooperative perception output example A
Figure 5. Representative cooperative perception output from the experimental pipeline.
Original project figure
Cooperative perception output example B
Figure 6. A second perception output showing multi-agent sensing and object-level results.
Interpretation

The comparison suggests that model selection, training settings, and sensing configuration should be evaluated for the intended operating environment. The reported loss reduction applies to this selected comparison; it does not establish performance across every scenario.

I extracted and organized evaluation data, then compared reliability across tested setups.

Using raw simulation files generated by other team members, I extracted key evaluation data and organized the results across sensor configurations and cooperative-perception schemes. I compared model accuracy and reliability across the tested setups, helping the team identify the most reliable hardware configuration among those evaluated.

This work strengthened my ability to turn raw simulation outputs into structured comparisons that could support system-design decisions.

The related TRB 2024 conference paper examines reliability in the specific setting of unprotected left turns.

Source: “场景驱动的自适应-自进化多车协同环境感知方法,” Tongji University, 23rd TongLuRen Transportation Technology Competition.

Next case studyIntelligent Train Driving Assistance →Explore another projectMetro Travel-Time Reliability →