Topics

#computer-vision #drive-by-wire #ros #self-driving-car #ssd #traffic-light-detection #udacity #udacity-self-driving-car #yolov3-tiny

Language breakdown

Python 95.3%

CMake 2.8%

C++ 1.8%

Dockerfile 0.1%

Shell 0.0%

▸ Files click to expand

README

Self-Driving Car using ROS

Udacity Self-Driving Car Engineer Nanodegree --- Capstone Project

Team 4Tzones

All team members contributed equally to the project.

4Tzones means "Four Time Zones" indicating that team members were located in 4 different time zones while working on this project. The time zones range from UTC+1 to UTC+8.

Software Architecture

Note that obstacle detection is not implemented for this project.

Traffic Light Detection Node

A large part of the project is to implement a traffic light detector/classifier that recognizes the color of nearest upcoming traffic light and publishes it to /waypoint_updater node so it can prepare the car to speed up or slow down accordingly. Because the real world images differ substantially from simulator images, we tried out different approaches for both. The approaches which worked best are described below.

OpenCV Approach

Simulator (Highway)

In this approach we used the basic features of OpenCV to solve the problem, the steps are described below.

Image is transformed to HSV colorspace, as the color feature can be extracted easily in this colorspace.
Mask is applied to isolate red pixels in the image.
Contour detection is performed on the masked image.
For each contour, area is checked, and, if it falls under the approximate area of traffic light,

polygon detection is performed and checked if the the number of sides is more than minimum required closed loop polygon.

If all the above conditions satisfy there is a red sign in the image.

Pros

This approach is very fast.
Uses minimum resources.

Cons

This is not robust enough, the thresholds need to be adjusted always.
Doesnt work properly on real world data as there is lot of noise.

SSD Approach (Single Shot Detection)

We need to solve both object detection - where in the image is the object, and object classification --- given detections on an image, classify traffic lights. While there are teams who approached it as 2 separate problems to be solved, recent advancements in Deep Learning has developed models that attempt to solve both at once. For example, SSD (Single Shot Multibox Detection) and YOLO (You Only Look Once).

Simulator (Highway)

Here we experimented with Tensorflow Object Detection API, using pretrained models on COCO dataset, such as: "ssdinceptionv2coco" and "ssdmobilenetv2coco":

Testing the coco pretrained models without retraining on simulator images didn't lead to any success,

since the simulated traffic lights look very different from real traffic lights, and hence we concluded, that if we were going to use this approach on the simulator, we would need a different model specifically retrained on the simulator images.

So we decided to utilize transfer learning, and retrain the models on images extracted from the simulator,

using 3 classes/labels only; Red, Yellow and Green.

We choose the "ssdinceptionv2_coco model", since it proved to be a good compromise between speed and accuracy,

and retrained it on the simulator images dataset provided by Alex Lechner here.

Sample dataset for simulator images simulator training images

The configuration parameters for retraining was:

* num_classes: 3. * fixedshaperesizer: 300x300, to reduce training time, since using larger image sizes during training didn't seem to increase the inference accuracy. * Dropout: True. * batch_size: 24. * num_steps: 20000, which experimentally proved to lead to good results.

The training took around 3.5 hours on an NVIDIA GTX 1070 (tensorflow-gpu == 1.4.0),

and the final training loss was around 2.x.

The retraining of the model lead to very good results; confidence levels reaching up to 0.999

even when the car is very far away from the traffic light:

Here are the results of our trained model.

simulator inference 1 simulator inference 2

Pros

This approach is very accurate.
It can detect all 3 colors; Red, Yellow & Green, with great confidence.
It can pinpoint the exact position and size of the lights, which can be further utilized

for accurately calculating the stopping line position.

Cons

It's slower than OpenCV method.

Real World (Test Lot)

We tested the pretrained models without retraining, on real world images from the ROS bags

provided by Udacity, which led to some success, since COCO dataset already has a Traffic Light class (No.10), however it was a limited success since the ROS bags images had unusual lighting; very bright in some cases, and often the 3 light colors were not distinguishable from one another and all looked somewhat yellow.

Similarly we opted for retraining the "ssdinceptionv2_coco" model, but this time we compiled

our own dataset, since datasets found online didn't lead to good enough results, so we labeled images from 3 different ROS bags provided by Udacity and added images from Bosch Small Traffic Lights Dataset here, which helped the model generalize better, and increased the detection confidence specially for instances when the traffic light was far away, since most images in the ROS bags have the traffic light in close proximity.

Here is a sample of the dataset. Udacity Test Site training images

The configuration parameters for retraining was:

* num_classes: 3. * fixedshaperesizer: 300x300. * Dropout: True. * batch_size: 24. * numsteps: 100000, here we increased the number of steps, since each step processesbatchsize images, so for example if we double the number of samples in the dataset, we will need to double the number of steps to achieve the same number of epochs, each epoch requires = (no. samples / batch_size) steps, and in this combined dataset we had around 22,000 samples/images.

The training took around 18 hours on an NVIDIA GTX 1070 (tensorflow-gpu == 1.4.0),

and the final training loss was around 1.x.

The results were good reaching to a confidence of 1.0 most of the time, but in some instances

the model completely fails specially when the traffic light is very close to the camera.

Here are the results of our trained model.

real inference 1 real inference 2

Pros

This approach is accurate in most cases.
It can detect all 3 colors; Red, Yellow & Green, with great confidence.
It can pinpoint the exact position and size of the lights, which can be further utilized

for accurately calculating the stopping line position.

Cons

It's not very fast, the FPS when running the ROS bag was averaging 15 FPS.
It requires a very large dataset including images of different lighting conditions,

different distances from the lights, etc, in order to be reliable.

YOLOv3-tiny (You Only Look Once)

Real World (Test Lot)

We used this approach for real world. TODO:write about it

Dataset

Image Collection

We used images from 3 ROS bags provided by Udacity:

Image Extraction

As described in How to export image and video data from a bag file, we:

<!--Replace <path-to-your-ros-bag> with the actual path to your ROS bag from which you want to extract images.-->
<!--Replace <topic> with the actual topic that contains images of your interest.-->
<launch>
  <node pkg="rosbag" type="play" name="rosbag" required="true" args="<path-to-your-ros-bag>"/>
  <node name="extract" pkg="imageview" type="extractimages" respawn="false" required="true" output="screen" cwd="ROS_HOME">
    <remap from="image" to="<topic>"/>
  </node>
</launch>

Prepared the environment by executing: roscd imageview && rosmake imageview --rosdep-install.
Created an extract-images-from-ros-bag.launch file (above).

- For traffic_lights.bag ROS bag we used /image_color topic. - For justtrafficlight.bag and loopwithtraffic_light.bag we used /image_raw topic.

Ran: roslaunch extract-images-from-ros-bag.launch.
Created a folder to keep extracted images in: mkdir <folder>.
Moved extracted images to the newly created folder: mv ~/.ros/frame*.jpg <folder>

Video Creation

We extracted images from the ROS bags in the Image Extraction step and converted them to videos following the instructions from How to export image and video data from a bag file. We:

Prepared the environment by executing: sudo apt install mjpegtools.
Ran:

ffmpeg -framerate 25 -i <folder>/frame%04d.jpg -c:v libx264 -profile:v high -crf 20 -pix_fmt yuv420p <output>, where <folder> is a directory with files extracted from a particular ROS bag and <output-name> is a desired name for your MP4 videos file (the file should have the .mp4 extension).

Below is a video archive containing 3 videos, each corresponding to one of the ROS bags mentioned in the Image Collection section. The archive is called "4Tzones Traffic Lights Videos" and is licensed under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

| | 4Tzones Traffic Lights Videos | |------------------------| :------------------------: | | Link | https://yadi.sk/d/DhyGqahR-NWtEA | | License | CC BY-SA 4.0 | | trafficlights.bag | trafficlights.mp4 | | justtrafficlight.bag | justtrafficlight.mp4 | | loopwithtrafficlight.bag | loopwithtrafficlight.mp4 |

Image Annotation

We used a Yolo_mark tool to label the extracted images. The annotated dataset which is called "4Tzones Traffic Lights Dataset" is available under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

| \ & \\ & \\\ | 4Tzones Traffic Lights Dataset | |------------------------| :------------------------: | | Link | https://yadi.sk/d/a1Kr8Wmg0zfa0A | | License | CC BY-SA 4.0 | | Total TL # of Samples | 2795 | | Red TL # of Samples | 682 | | Yellow TL # of Samples | 267 | | Green TL # of Samples | 783 | | No TL # of Samples | 1063 |

\* TL stands for "Traffic Lights" and # stands for "Number."

\\ Notice that the total number of images contained in the ROS bags mentioned above is a little bigger. We removed all images that are ambiguous, e.g., two traffic light bulbs are simultaneously ON, or the image border partially cuts a traffic light.

\\\* It takes about 3 hours of continuous work for one person to label images from all three ROS bags using Yolo_mark given that he has a decent monitor and mouse.

Image Augmentation

We tried different neural networks for traffic lights detection and classification. We first used the data obtained during the Image Annotation step. Models trained on these data did not perform well enough on similar but previously unseen images. The 4Tzones Traffic Lights Dataset is just not good enough to enable the neural network to generalize. The dataset is unbalanced in different aspects. That is, the number of samples per class is not equal to each other; the majority of red traffic light images are captured from far distances; in all the traffic light images containing a close view of the traffic light, the traffic light position is biased to the left, and in other aspects. After several trial and error attempts, it was obvious that we need to augment the dataset. Moreover, for different models we used different training code, that is, for YOLO-tiny model we used code from the keras-yolo3 repository with minor modifications and for SSD models we used the TensorFlow Object Detection API repository. Both training scripts accept different labels format, and we needed to convert Yolo_mark annotations to those other formats. To accomplish these image augmentation and label conversion tasks we have created a data_preparer.py. With this script, we easily augmented the 4Tzones Traffic Lights Dataset, which dramatically increased the models' ability to generalize. More about the data_preparer.py script is in the Data Preparer Script section. The resulting annotated augmented dataset, which is called "4Tzones Traffic Lights Augmented Dataset" can be downloaded using the link below and is licensed under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

| \ & \\* | 4Tzones Traffic Lights Augmented Dataset | |------------------------| :------------------------: | | Link | https://yadi.sk/d/q2Yyy9PO2SrMKQ | | License | CC BY-SA 4.0 | | Total TL # of Samples | 11988 | | Red TL # of Samples | 1998 | | Yellow TL # of Samples | 1998 | | Green TL # of Samples | 1998 | | No TL # of Samples | 5994 |

\* Notice that the number of samples of red, yellow, and green traffic lights is equal to each other and the number of samples of images without traffic lights is triple that. Such a balancing of the dataset is suggested in the How to improve object detection section of README file from https://github.com/AlexeyAB/darknet repository by AlexeyAB.

\** Notice that you cannot obtain the same level of "balancing" by only using the data_preparer.py script on the 4Tzones Traffic Lights Dataset. The process of creating the 4Tzones Traffic Lights Augmented Dataset involved several manual steps, that can be described as follows. Initially, we had stored labeled images from different ROS bags in different folders. On each dataset, we performed flipping, scaling, and balancing. For balancing, the number of samples per class was set to about 2000 (the data_preparer.py script has a --balance N option which commands the script to create N sample images per class through augmentation and 3*N samples for images without traffic lights). It was needed to balance the number of samples among the datasets. That is, suppose that we have N1 red traffic light samples from the first ROS bag, N2, and N3 from the second and third ROS bags. If we combine these three datasets and then balance the combined dataset with samples per class parameter set to 1998 (i.e. --balance 1998), we would get 1998 red traffic light images with a proportion of images from different datasets equal to N1:N2:N3. To avoid such an uneven proportion, we would need to generate a sufficient amount of red traffic light images for each dataset (2000 samples, for example) and then pick 666 images from these 2000 generated samples in accordance with the uniform distribution. Then, combining 666 images from each of the three datasets, we would get 1998 samples of red traffic light images in the final dataset with a proportion of the presence of images from each of the three initial datasets equal to 666:666:666. For images without traffic lights, this proportion would be equal to 1998:1998:1998. So, we performed such a "fair" balancing among the three datasets for red, yellow, green, and images without traffic lights and then combined them into one dataset called the 4Tzones Traffic Lights Augmented Dataset.

Data Preparer Script

For data augmentation and conversion of labels to different formats we have created a data_preparer.py script. It is a quite sophisticated script that is capable of performing horizontal image flipping, image scaling, adjustment of brightness and contrast, image resizing, dataset balancing, that is, making number of samples of red, yellow, and green lights equal to a specified value while creating triple that for samples without traffic lights, random picking of the specified number of samples, and conversion of image annotations (bounding boxes) to several different label formats, such as a format used in the Bosh Small Traffic Lights Dataset, a format produced by Yolo_mark tool, a format required by keras-yolo3, a format used in the Vatsal Srivastava's Traffic Lights Dataset. The data_preparer.py --help command produces a help-message that presents comprehensive instructions on how to use the script. We strongly recommend reading it before feeding your data to the script.

usage: data_preparer.py [-h] --dataset {boschsmalltrafficlights,vatsalsrivastavatrafficlights,yolo_mark} [--fliplr] [--scale] [--balance [B]] [--pick N] [--resize H W] --input-dir DIR --output-dir DIR [--continue-output-dir] [--draw-bounding-boxes] This script is capable of working with several datasets from the list below. It applies the requested image augmentation to the images from the provided dataset and converts labels to several formats specified below. It also balances dataset to the following form: red == yellow == green == nolight/3. Datasets: - Bosh Small Traffic Lights Dataset: https://hci.iwr.uni-heidelberg.de/node/6132 - Vatsal Srivastava's Traffic Lights Dataset (Simulator & Test Lot): https://drive.google.com/file/d/0B-Eiyn-CUQtxdUZWMkFfQzdObUE/view?usp=sharing - Any Traffic Lights Dataset Labeled with Yolomark: https://github.com/AlexeyAB/Yolomark. 4Tzones Traffic Lights Dataset (Yolo_mark compatible): https://yadi.sk/d/a1Kr8Wmg0zfa0A. Label formats: - One row for one image (singular and ternary); Useful for https://github.com/qqwweee/keras-yolo3; Row format: imagefilepath box1 box2 ... boxN; Box format: xmin,ymin,xmax,ymax,class_id (no space). - Vatsal Srivastava's yaml format (only ternary). Example: - annotations: - {class: Green, xwidth: 17, xmin: 298, yheight: 49, ymin: 153} class: image filename: ./images/a0a05c4e-b2be-4a85-aebd-93f0e78ff3b7.jpg - annotations: - {class: Yellow, xwidth: 15, xmin: 364, yheight: 43, ymin: 156} - {class: Yellow, xwidth: 15, xmin: 151, yheight: 52, ymin: 100} class: image filename: ./images/ccbd292c-89cb-4e8b-a671-47b57ebb672b.jpg - Bosh Small Traffic Lights yaml format (only ternary). Example: - boxes: - {label: Red, occluded: false, xmax: 640, xmin: 633, ymax: 355, ymin: 344} - {label: Yellow, occluded: false, xmax: 659, xmin: 651, ymax: 366, ymin: 353} path: ./images/ccbd292c-89cb-4e8b-a671-47b57ebb672b.png - Yolomark format. One file per image. Example: imagename.jpg -> image_name.txt. Content: <object-class> <xcenter> <ycenter> <width> <height> <object-class> <xcenter> <ycenter> <width> <height> ...

optional arguments: -h, --help show this help message and exit --dataset {boschsmalltrafficlights,vatsalsrivastavatrafficlights,yolo_mark} dataset name --fliplr apply imgaug.Fliplr function (flip horizontally) to all images; dataset size will x2 in size --scale apply imgaug.Affine(scale=0.7) function (scale image, keeping original image shape); dataset size will x2 in size --balance [B] balance dataset, so that there is an equal number of representatives of each class; when no argument is provided, the number of elements per RED, YELLOW, GREEN classes are made equal to the maximum number of elements per class after the first processing stage, i.e., before balancing; if B argument is provided, the number of samples per RED, YELLOW, and GREEN classes are made equal to B; number of instances for NO_LIGHT class is made equal to 3*B --pick N picks N images from the original dataset in accordance with uniform distribution and ignores other images --resize H W resize all images to the specified height and width; aspect ratio is not preserved --input-dir DIR dataset's root directory --output-dir DIR directory to store prepared images and labels --continue-output-dir expand existing output directory with new image-label entries --draw-bounding-boxes draw bounding boxes on the output images; do not use it while preparing data for training

Other approaches for traffic light detection

We experimented with few other (unsuccessful) approaches to detect traffic lights.

Idea

The idea is to use the entire image with a given traffic light color as an individual class. This means we will have 4 classes

1. Entire image showing yellow traffic sign 2. Entire image showing green traffic sign 3. Entire image showing red traffic sign 4. Entire image showing no traffic sign

Models

We trained couple of models:

A simple CNN with two convolutional layers, a fully connected layer and an output layer. The initial results looked promising with training accuracy > 97% and test accuracy > 90%. However when we deployed and tested the model, the results were not consistent. The car did not always stop at red lights and sometimes it did not move even when the lights were green. Efforts to achieve higher accuracies were in vain.

Used transfer learning for multi-class classification approach using VGG19 and InceptionV3 models, using imagenet weights. The network did not learn anything after 1-2 epochs and hence the training accuracy never exceeded 65%.

Learning Points

Future Work

Acknowledgements

We would like to thank Udacity for providing the instructional videos and learning resources.
We would like to thank Alex Lechner for his wonderful tutorial on how to do transfer learning on TensorFlow Object Detection API research models and get it to run on older tensorflow versions, as well as providing datasets. You can view his readme here: https://github.com/alex-lechner/Traffic-Light-Classification/blob/master/README.md#1-the-lazy-approach

🔗 More in this category