Track: Programmable Real-Time Networks and Applications |
| Enhancing Autonomous Vehicle Perception with Synthetic Data: Tackling Class Imbalances in Critical Scenarios |
| The adoption of autonomous vehicles (AVs) depends heavily on robust perception systems trained on diverse datasets. However, a significant challenge is the scarcity of rare yet critical classes like ambulances in publicly available datasets. This class imbalance impacts the ability of AVs to detect and respond to emergency vehicles accurately, posing safety risks in high-stakes scenarios. Synthetic data generation effectively addresses this limitation by leveraging advanced simulation platforms such as CARLA, LGSVL, and AirSim. These simulation tools allow the creation of highly customizable datasets with rare classes in diverse conditions, including varying lighting, weather, and urban environments. By generating synthetic scenarios tailored to underrepresented objects like ambulances, we can enhance training datasets without the constraints of real-world data collection. This approach improves object detection accuracy and performance in edge cases critical to AV safety. In this session, we will explore how synthetic data generation can bridge the gap left by public datasets and provide practical insights into integrating simulation-based data into AV development pipelines. We will focus on the importance of balancing datasets for classes like ambulances and demonstrate the impact of this augmentation on model performance. Attendees will understand how synthetic data improves AV reliability, paving the way for safer and more efficient autonomous driving systems. This discussion underscores the transformative role of simulation in addressing real-world challenges in AV perception. |
|
| Presentation Video |
| Presentation Notes |
|
Jaiswal_Enhancing-Autonomous-Vehicle-Perception-with-Synthetic-Data-Tackling-Class-Imbalances-in-Critical-Scenarios_V1.0.pptx |
Real Time Communications Conference & Expo at Illinois Tech
IEEE International Conference
Follow Us
Share This
