_____ _ _____ _ _
/ ____| (_) / ____| | | | |
| | _ __ _ _ _ ___ ___ | | ___ _ __ | |_ _ __ ___ | |
| | | '__| | | | / __|/ _ \ | | / _ \| '_ \| __| '__/ _ \| |
| |____| | | |_| | \__ \ __/ | |___| (_) | | | | |_| | | (_) | |
\_____|_| \__,_|_|___/\___| \_____\___/|_| |_|\__|_| \___/|_|_
Garrett Pitcher garrett.pitcher@gmail.com
Mitch Maifeld udacity@maifeld.name
Hanqiu Jiang hanq.jiang@gmail.com
Shaun Cosgrove shaun.cosgrove@bogglingtech.com
Wei Guo guoweist@foxmail.com
The basis of the traffic light classification is a Deep Neural Network Single Shot MultiBox Detector (SSD) architecture. This classifies the object and detects a bounding box for said object. For greater classification accuracy, we created two models, one trained on simulator images and one trained on images taken from a run around the course on Carla (extracted from a rosbag provided by Udacity). These images were extracted and labeled using the VIA tool (see below). A pre-trained model from Tensorflow detection model zoo was re-trained with the Tensorflow Object Detection API. The underlying MobileNet architecture trades accuracy for high efficiency and enables us to detect traffic light with high fps.
Due to a lack of suitable Machine Learning acceleration hardware available among the development team, we chose to process a reduced number of video frames for the traffic light classification. The planner is crafted with this limitation in mind and the end result is a dependable traffic light detection mechanism that strikes a balance between accuracy and performance.
Detection in simulator:
Detection in rosbags:
The Planner takes the high-level information about the world, information about the current and past states of the vehicle and what the vehicle currently perceives, meshes this together and generates both a position and velocity track for the vehicle to follow. Based on the current position of the vehicle, it selects a set of target world way points to visit and using a spline fitting utility, creates intermediate target way points (with associated velocities) for the vehicle to follow. This will ensure that the vehicle is always tracking to a smooth path and will re-track to the center of the lane in case it ever goes off center of the target. The design can be easily extended to enable lane keeping, lane changing or obstacle avoidance; This means operation in a multi-lane environment with other traffic. Traffic lights are processed by determining which set of lights the vehicle needs to obey next and 'looking' at the state of the lights as the vehicle approaches them. Based on the dynamic interaction of the vehicle and the traffic lights, the autopilot assesses the intersection state and chooses an appropriate behavior - prepare to stop, stop or proceed.
RVIZ was used to visualize the generated way-points from the planner.
A PID controller is used for the Throttle and a PD controller is used for the Brake. We began tuning by using the on/off relay method to determine time periods to maximum desired speed and declerations to a stop. We used these as inputs to the Zigler-Nichols PID ratios to initialize our tuning. From there, we manually refined the proportional gain, integral, and derivative parameters to reach the response shown in the figure below. We created the plot in rqt_plot to show the throttle PID's recovery stability. The blue line shows our commanded throttle setting; the red line shows the vehicles response to the command. Prior to 282 in the figure, the vehicle was manually controlled to an over-desired velocity. Removing the manual control at 282, the automatic throttle kept a zero until the vehicle coasted back to the desired speed beginning about 289 and reaching steady state at 291. At 297, we switched to manual again and applied the brakes to reach a stop. At 300, we return to automatic control where the vehicle applies full throttle (such as at a stop light) to clear the intersection and quickly reach desired speed. It reaches that desired speed near 315 with minimal overshoot but quick return to set point.
Cubic Spline library by Atsushi Sakai
An extremely approachable aritcle, Racoon Training by Dat Tran was used to guide the initial steps into the tensorflow object detection utilities. Tools from his repository were used or modified to create the protobuf binaries for training the object detector.