Скачать или смотреть NXP AIM Challenge Final Run -Team DefianzRacing

Sources:
GitHub Repo: https://github.com/TDR-SDC/NXP_Contro...
Original Stream Link:    • Smart Car Race Design Challenge 2021 - Gra...  
Watch From 02:47:00

Competition Website: https://nxpaimindia.com/


I am very proud to announce that our we have secured 3rd Position in "NXP Artificial Intelligence in Mobility(AIM) Challenge" organized by NXP Semiconductors, NXP AIM India. In Addition we were also felicitated with "Best Car Model Award" for creating the most attractive car design.

The Challenge required us to create a Self-Driving Car model which was simulated in Gazebo using ROS2. Our model had to be robust enough to identify the area of track; drive without crossing the track limits while avoid static and dynamic obstacles such as pedestrians, cars and barricades; And at the same time obey the traffic signs(stop, left turn, right turn) and traffic light.

This project required every bit of knowledge we had about self driving technology combined with exceptional hard work and out-of-the-box thinking. Even though most other teams used LIDAR for obstacle avoidance we took an unconventional approach to only use the camera for driving, thus following a Humanistic approach for driving.

We created and trained a custom CNN model in TensorFlow for controlling our car it uses the images from onboard camera to identify obstacles and give steering inputs while respecting track boundaries. We created it from scratch and trained it in increments with the track getting more complex after appreciable loss(mse) is achieved. and used the best weights for next training stage with more obstacles, thus using the concept of genetic learning in supervised learning.
We used YOLO V4 tiny for traffic sign Recognition to follow traffic rules, using transfer learning from GTSDB (German Traffic Sign Detection Benchmark) dataset to achieve the highest possible accuracy.