TensorRT deployment for Birds Eye View Representation in AVs

Deployed a Bird’s Eye View Representation algorithm for self-driving cars in Jetson Xavier aarch64 machine. The algorithm provides road segmentation in the birds’ eye view and involves EfficientNet-b0 feature extraction model for encoding the camera and lidar input from Nu-Scenes.

Accelerated the algorithm by exporting the EfficientNet feature extraction model and ResNet model into TensorRT engine via Onnx to achieve 15x speedup.

Sample code

Topic Covered:

1. Basics of Onnx and TensorRT.
2. Datasets- Kitti, Nuscenes (with Lidar).
3. Experience with Jetson devices

Programming Language: Python, Pytorch