The Indoor Positioning System detects the position, orientation and identity of multiple vehicles simultaneously. The position and orientation are updated at 50 Hz.
The Indoor Positioning System works with active light sources (LEDs) on the vehicles and a Camera looking down from the ceiling. The LEDs are detected based on their high brightness. The camera is set to a very short exposure (~100 microseconds). Thus, the ambient light creates a very small signal (almost black), while the LEDs still appear as white dots. The short exposure time also eliminates the problem of motion blur. At a top speed of 4 m/s the vehicle travels 0.4 mm during the exposure.
Other light sources and reflective surfaces on the vehicle can create false signals and must be covered with tape. This includes the connectors on the Raspberry Pi and the status LEDs on the motor speed controller.
The outer three LEDs indicate the vehicle pose and are permanently illuminated. The section Pose Calibration describes how the LED positions are related to the reference pose.
The central LED flashes in a pattern that is different for each vehicle. The patterns are chosen such that sampling effects do not create ambiguous signals. TODO reference MA Isabelle Tuelleners, cf. Nyquist–Shannon sampling theorem.
|Period of flashing, in number of frames||4||7||10||13||16||7||10||13||16||19||10||13||16||19||22||13||16||19||22||25||16||19||22||25||28|
|LED illumination duration, in number of frames||2||2||2||2||2||5||5||5||5||5||8||8||8||8||8||11||11||11||11||11||14||14||14||14||14|
The IPS software has two major components, the LED detection and the IPS pipeline. The LED detection reads images from the camera at 50 Hz and extracts the list of image coordinates for all visible LEDs. These
LedPoints are published via DDS. The raw images are not saved or published, as this would create impractically large data volumes. The IPS pipeline processes the LED points and publishes
VehicleObservations, which include the vehicle's position, orientation, identity and a timestamp.
The LED detection uses the OpenCV functions
cv::moments to find the centers of the LEDs.
The IPS pipeline performs a relatively complex data processing task. To make the software easier to understand, it is broken down into independent, stateless processing steps.
The UndistortPoints function transforms the image/pixel coordinates to metric coordinates on the driving surface (floor). The calculation is based on a calibration image. See https://git.rwth-aachen.de/CPM/Project/Lab/software/tree/master/matlab_scripts/camera_calibration
The DetectVehicles function groups and classifies points as
back_right. Points that appear to not belong to a vehicle are discarded.
The Queue collects the 50 most recent samples of the vehicle points. This is necessary, as the vehicle ID can not be determined from a single image.
The DetectVehicleID function extracts the ID for each vehicle from the last second (50 frames) of vehicle points.
The PoseCalculation function calculates the vehicle reference pose based on the
back_right points. The calculation is based on manually collected calibration data. See https://git.rwth-aachen.de/CPM/Project/Lab/software/tree/master/matlab_scripts/ips_pose_calibration
The implementation can be found at https://git.rwth-aachen.de/CPM/Project/Lab/software/tree/master/ips2/src
First, other PVC canvases are removed from the floor to reveal the checkerboard canvas. Then an image of the checkerboard pattern is taken using the
PylonViewerApp. The exposure is adjusted such that the checkerboard corners are well resolved. If the exposure is too short or long, the corners are "rounded off", which degrades the quality of the calibration.
The Matlab function
detectCheckerboardPoints is used to determine a list of matching points
(x_floor, y_floor, x_image, y_image).
TODO explain calibration jig