Page 177 - AI for Good-Innovate for Impact Final Report 2024

AI for Good-Innovate for Impact

(continued)

Domain Transportation
Metadata (type Images and labels with structured standard
of data) 40-GEELY
Model Training • Vehicle vision perception model for parameter optimization
and fine-tun- • Based on style-transfer method model fine-tuning.
ing

Case Studies China, MIIT.gov, intelligent connected vehicle test management procedure
and industrial development suggestion

Testbeds or link
pilot deploy-
ments

40�2� Use case description

40�2�1 Description

Introduction: The level of autonomous driving at this juncture does not yet align with the
completion objective of Level 4 unmanned operation, primarily attributed to the imperative
of algorithmic enhancement, and a concomitant deficiency of adequate algorithmic data.
Autonomous driving deployments necessitate an abundance of real-world data to facilitate
training and testing processes, but the procurement and annotation of this data is inevitably
expensive and labor intensive.

Furthermore, the acquired data may exhibit biases, and fail to encompass every conceivable
driving scenario. Autonomous driving systems are bound by the requirement to withstand
testing under diverse and complex circumstances, such as harsh weather situations, congested
traffic conditions, and ongoing road construction projects. Nevertheless, these specific
conditions may not be consistently accessible for real-world testing.

In the dimension of validation and testing, the restrictions of closed testing arenas include a
monolithic environmental context, an inability to replicate varying complexities encountered on
real-world roads, a deficiency in affirming performance on actual roadways, and an exorbitantly
high cost associated with utilizing testing equipment and spaces. Testing processes undertaken
on actual roads are dependent on real road and traffic situations, which mandate considerable
investment of manpower and resources, whilst also introducing safety risks.

The case in study utilizes virtually simulated synthetic data to amplify the perceptual capabilities
of the autonomous driving system, and also to augment the efficiency of testing and validation
processes. This avant-garde approach plays a pivotal role in elevating the system's performance,
safety, and adaptability, while concurrently accelerating the progression of system development
and testing. This consequently results in a substantial reduction in costs and risks.

Tool Overview: This case is applied to intelligent driving and intelligent transportation
scenarios. The virtual simulation platform is used to build scenarios, generate a large number
of data under various scenarios, including normal and extreme situations, and simulate various
complex driving scenarios, such as bad weather, traffic congestion, road engineering, etc., to
generate more diversified data. Helps intelligent driving systems better generalize to previously

161