Module for detecting traffic lights in the CARLA autonomous driving simulator. Based on the YOLO v2 deep learning object detection model and implemented in keras, using the tensorflow backend.
⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️
This repo is deprecated. For a newer and improved model, implemented in PyTorch, please refer to this repo.
⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️
Module for detecting traffic lights in the CARLA autonomous driving simulator (version: 0.8.4).
Built upon and inspired by https://github.com/experiencor/keras-yolo2.
Instructions and more traffic light detection examples can be found below.
This module is used along several other modules to implement our version of imitation learning in the CARLA simulator. Results of the core module can be found on this repository
Model for objection detection is based on tiny yolov2
Training started with yolov2 coco pretrained weights
It was first trained on the LISA traffic light detection dataset (~5800 images), and after that on the dataset collected from the CARLA simulator by myself (~1800 images).
Dataset collected by myself in the CARLA simulator can be found here, annotations can be found here.
Important note - several images in the dataset are left out of annotations because bounding boxes are too small (too far away). I also filtered (left out) all images that have xmax < 15 when loading the dataset. There is around 70-80 out of ~1800 images that are left out, so it isn't that problematic.
Pretrained model can be found here.
To train:
To evaluate:
To generate anchors:
Soon to be added: