Tesla 따라하기
Project Goal
- Reduce Memory Consumption
- Develop a Complete End-to-End Pipeline
- Enable Real-Time Autonomous Control
Project Page
Detailed information about the project can be found in the project page above!
Project Overview
I worked on a project that integrated object detection, lane detection, depth estimation, and planning algorithms to develop an autonomous driving system. Through the perception stage, we processed the obtained information to generate an occupancy map and planned the control process accordingly.
For object detection, we used YOLOv8, while UFLD was used for lane detection, and Metric3D for depth estimation. In the planning stage, we implemented path generation using the A* algorithm.
To improve inference speed, we converted the PyTorch models to TensorRT. After developing the object, lane, depth, and planning engines, we integrated them into a single unified engine. We named the unified engine as the Autonomous Engine.
For our project, we used the Nvidia Jetson Orin Nano 8GB, Nvidia JetRacer, and a CSI camera. While driving the JetRacer, we collected a total of 11,243 images. After capturing the data, we manually labeled objects and lanes before conducting experiments.
Results
This is a visualization of the results when running the Autonomous Engine.
During this project, the vehicle had difficulty accurately detecting lanes, so we used ResNet to train the direction it should follow. This video visualizes the results.
Seminar
I was in charge of the presentation at the 10th deepdaiv Open Seminar.😊
autonomous-engine