Page 69 - AI for Good-Innovate for Impact
P. 69
AI for Good-Innovate for Impact
Use case –13: Intelligent Automobile Integrated Safety and Secure
Service Platform 13-Changan
Country: China
Organization: Chongqing Changan Automobile
Contact person: Dr. Yonggang Luo, luoyg3@ changan .com .cn
13�1� Use case summary table
Domain Transportation
The Problem to be Enhance the safety and security level of autonomous driving, reduce
addressed traffic accidents, ensure the safety of life, and increase the degree of
traffic automation.
Key aspects of the Integrated perception-decision closed-loop simulation testing
solution platform, achieving closed-loop testing in all scenarios through
generative and reinforcement methods.
Technology keywords Autonomous driving, simulation, perception-decision safety, feder-
ated learning, data barrier
Data availability Partially available.
Metadata (type of POI (point of interest) data; Object detection data; BEV images data;
data) Adversarial patches data; trajectory data
Pipeline Data collection and pre-processing; scenario generation; inference;
evaluation; Feedback loop for continuous improvement
Case Studies Analysis of safety of perception-decision system through predictive
simulations.
Testbeds or pilot Virtual simulations in controlled environments; Real-vehicle adversar-
deployments ial attack Tests; Real-world pilot projects in selected traffic scenarios;
Ablation comparison test.
Metrics, KPIs, Reduction in accident rates; Adversarial efficiency compared to natu-
measurements ral data; Data utilization rate; Perception attack success rate.
GPU Internal GPUs.
Network require- 4g/5g network for collaborative training/testing, cloud service, auton-
ments, architecture omous driving simulator
components
Role of Trainings, Developed the platform based on IEEE SA P3129, P3187, CSAE
standards standards and discussed related technical topics with experts in
workshops (e.g DIAVT 2024)
Pre-standard research Benchmarks on advanced autonomous driving simulation and testing
algorithms & tools; Benchmarks on cross domain federated learning;
53
