Carla Self Driving Car System Integration Project

Team Dragon

Name	Email
Waleed Mansoor	[email protected]
Jaewoo Park	[email protected]
Sasha Jaksic	[email protected]
Dominik Marquardt	[email protected]
Juil O	[email protected]

System Architecture

This system has three main parts.

Perception

The main role of this part is detecting the traffic light and publishing '/traffic_waypoint' topic so that the next planning part can generate an appropiate trajectory.

For this, we trained the pre-trained classifier using some dataset. The pre-trained graph is ssd_mobilenet_v1_coco

We used images from ROS topics of simulator and bag files as training dataset. We can see these images following commmands.

roslaunch launch/styx.launch
rosrun image_view image_view _sec_pre_frame:=0.1 image:=/image_color

rosbag play just_traffic_light.bag
rosrun image_view image_view image:=/image_raw

And then that can be saved with this command.

rosrun image_view image_saver _sec_per_frame:=0.1 image:=/image_raw

The waypoint to be stopped at when the light is red is provided by stop_line_positions config. So the classifier tries to find a traffic light when the car comes near the closests stop_line.

# find the closest visible traffic light (if one exists)
min_dist = 100000
for stop_line_wp_idx in self.stop_line_wp_idxs:
    dist = stop_line_wp_idx - self.car_wp_idx
    if dist >= 0 and dist < min_dist:
        min_dist = dist
        if min_dist < self.visible_distance_wps:
            # It uses the stop_line postion rather than the traffic light position
            light_wp = stop_line_wp_idx

# If there is a visible traffic light
if light_wp != -1:
    # if camera is on: through classifier
    if self.has_image:
        state = self.get_light_state(light_wp)
        return light_wp, state
    # if camera is off: through ground truth
    else:

Planning

Control

Installation

It is probably easiest to install ROS and deploy everything with Docker, and although you could use the VM provided by Udacity here, it is recommended to deploy using Docker.

Clone the repository on your host OS to your PWD

cd $PWD/Dragon

If you do not have Docker, feel free to install Docker for your host OS. Install Docker

Build the docker container

docker build . -t capstone

Run the docker file

docker run -p 4567:4567 -v $PWD:/Dragon -v /tmp/log:/root/.ros/ --rm -it capstone

Source the ROS env variables. (Change Kinematic to Indigo if using 14.04)

echo "source /opt/ros/kinetic/setup.bash" >> ~/.bashrc
echo "source /Dragon/ros/devel/setup.bash" >> ~/.bashrc

Clone this project repository

Install python dependencies

cd /Dragon
pip install -r requirements.txt

Make and run nodes

cd ros
catkin_make
source devel/setup.sh

At this point, exit bash on Docker or proceed to Run steps below;

How to "SSH" into your Docker container

Note that for every new bash session if you are SSH-ing into the Docker container, you should open a new session and use the docker exec command.

docker ps

You can choose to grep the container ID or name and pipe it into the exec command directly or just use the clipboard to copy and paste

docker exec -it <container name or id> /bin/bash
cd /Dragon/ros

This will successfully get you a new bash window to the container where your ros environment resides.

---

Run ROS alongside the simulator

After intalling successfully, and run below respectively.

roscore

roslaunch launch/styx.launch

---

Running the Traffic Light detection standalone

1. Open 2 different bash shells
2. In both bash windows do the following. (If you have installed with the instructions above, skip this step.)

cd Dragon/ros
source devel/setup.bash

3. In window 1,

roslaunch launch/styx.launch

4. In window 2,

roslaunch tl_detector tl_detector.launch

---

How to test dbw

In order to test dbw, you need to download the ros.bag file and save it to the $PWD/Dragon/data directory

cd /Dragon/data
curl -H "Authorization: Bearer YYYYY” https://www.googleapis.com/drive/v3/files/0B2_h37bMVw3iT0ZEdlF4N01QbHc?alt=media -o udacity_succesful_light_detection.bag
mv udacity_succesful_light_detection.bag dbw_test.rosbag.bag

After downloading and renaming the file,

cd ../ros
roslaunch twist_controller dbw_test.launch

This will save 3 csv files which you can process to figure out how your DBW node is performing on various commands.

/actual/* are commands from the recorded bag while /vehicle/* are the output of your node.