Autonomy Intern @ Oshkosh

Semantic Segmentation for Autonomous Vehicles

During my internship at Oshkosh Corporation’s Autonomy Team, I focused on developing and deploying semantic segmentation models for advanced perception in autonomous vehicles. This project involved creating robust segmentation systems capable of identifying road elements, vehicles, pedestrians, and other critical objects in real-time. Scripted ROS2 nodes to integrate the model in the stack, performance was tested with NVIDIA Orin device.

Key Contributions

• Model Development:
  Implemented and fine-tuned deep learning segmentation models for real-time object detection and segmentation.
• Optimization:
  Implemented TensorRT acceleration for deployment on NVIDIA Orin platforms, reducing inference latency by 30%.
• Integration:
  Integrated the segmentation module into the existing Autonomy stack for enhanced autonomous navigation and tested on vehicle.

Challenges and Achievements

• Challenge
  Optimizing segmentation models to handle dynamic urban environments (e.g., moving vehicles, changing lighting conditions).
• Achievement
  Achieved >60% mIoU (mean Intersection over Union) on the custom dataset with a lightweight architecture, meeting real-time performance requirements.

Impact

The developed segmentation system significantly improved object detection accuracy and path planning reliability for autonomous vehicles. This work contributed to Oshkosh’s ongoing efforts in advancing autonomous vehicle technology for various applications.

Demo Video

`