Luis Mejias, Srikanth Saripalli, Pascual Campoy and Gaurav Sukhatme
Introduction Related work Testbed Visual preprocessing Control Architectures Experiments Conclusion
Introduction Related work Testbed Visual preprocessing Control Architectures Experiments Conclusion
Goal: vision-guided autonomous flying robots Application: Law enforcement, search and rescue, inspection and surveillance Technique: Object detection, tracking, inertial navigation, GPS and nonlinear system modeling
In this paper: Two UAVs – Avatar and COLIBRI Visual tracking => control commands
Introduction Related work Testbed Visual preprocessing Control Architectures Experiments Conclusion
Hummingbird (A. Conway, 1995) Model-scale Use GPS only 4 GPS antennas Precisions: position 1cm attitude 1 degree
AVATAR (Jun, 1999) Onboard INS & GPS Kalman Filter for State Estimation Simulation
Vision-guided Helicopter (Amidi, 1996, 1997) Onboard DSP-based vision processor Combine GPS and IMU data
Vision-augmented navigation system (Bosse, 1997) Uses vision in-the-loop to control a helicopter Visual odometer (Amidi, 1998) A notable vision-based technique used in autonomous helicopter (Wu, et al, 2005) Vision is used as additional sensor and fused with inertial and heading measurements for control
Introduction Related work Testbed Visual preprocessing Control Architectures Experiments Conclusion
AVATAR Gas-powered radio-controlled model helicopter RT-2 DGPS system provides positional accuracy of 2 cm ISIS-IMU provides rate information to onboard computer, which is fused using a 16 state Kalman filter Ground station: a laptop to send high-level control commands and differential GPS corrections Autonomous flight is achieved using a behavior-based control architecture
COLIBRI Gas powered model helicopter Fitted with a Xscale based flight computer augmented with GPS, IMU, Magnetometer, fused with a Kalman filter VIA mini-ITX 1.25 GHz computer onboard with 512 Mb RAM, wireless interface and a firewire color camera Ground station: a laptop to send high-level control commands, and for visualization
Introduction Related work Testbed Visual preprocessing Control Architectures Experiments Conclusion
Image segmentation and thresholding Convert the image to grayscale Use the value of “target color” as threshold Segment the image to binary image where the object of interest is represented by 1’s and background with 0’s
Square Finding Find contours (represented by polylines) from the binary image Use an algorithm to reduce the points in polylines Result: simplified squares
Template Matching User selects a detected window (a target)from the GUI A patch is selected around the location of the target Use local search window to find best match between the target and the detected contours, deciding which window to track
Kalman Filter Once a suitable match is found, a Kalman filter is used to track the feature positions Input: x and y coordinates of the features Output: estimates of these coordinates in the next frame
The user selects the object of interest from the GUI The location of the object is used to generate visual reference
Lateral visual reference
Vertical Visual Reference
Introduction Related work Testbed Visual preprocessing Control Architectures Experiments Conclusion
A hierarchical behavior based control architecture Output of Kalman filter is compared with desired values to give an error signal to controller
Controller is based on a decoupled PID control
Introduction Related work Testbed Visual preprocessing Control Architectures Experiments Conclusion
At Del Valle Urban Search and Rescue Training site in Santa Clarita, California AVATAR, four trials First, the helicopter is commanded to fly autonomously to a given GPS waypoint As soon as it detects the featured window, the controller switches from GPS-based to vision-based control
Location of the features in the image
Helicopter position in meters. (left figure) vertical axis– easting (right figure) vertical axis – northing
At ETSII Campus in Madrid, Spain COLIBRI Seven experimental trials on two different days
Velocity references (vy r ) with the helicopter velocity (vy) Lateral displacement (east)
Velocity references (vz r ) with the helicopter velocity (vz) altitude displacement (down)
Helicopter displacements during the entire flight trial
colibrivideoWeb.wmv colibrivideoWeb.wmv
Introduction Related work Testbed Visual preprocessing Control Architectures Experiments Conclusion
Demonstrated an approach to visually control an autonomous helicopter: use visual algorithm to command UAV when GPS has dropouts Experimentally demonstrated by performing vision-based window tracking tasks on two different platforms at different locations and different conditions
The topic is interesting Visual algorithm is demonstrated effective in the experiments But… the writing is so ugly. Poor explanation ▪ features, template and matching Incomplete explanation of figures