Tracking noise, 0 to 4 px; the disturbance the thesis names. In Auto the model sets the parameter and runs the cases of the paper; in Hold the slider sets it. The line below the control says what the model does at this moment. With reduced motion the model shows a fixed frame, and the controls still work.
Lab · Live model · 06
The coast in the camera.
The image is the one in which the controller of the thesis operates. The detection gives the bounding box of the coastline, its four corners are the image features, and their desired positions move along the frame so that the vehicle advances along the coast. The camera holds the velocity of the last solution until the measured features depart from the features that the solution predicted.
What the drawing shows
The frame is the camera of the thesis, 720 by 480, with the marks of a viewfinder. The coastline runs across it with the sea below, and the faint line beyond the frame is the coast the camera does not see. In the image of the thesis the coast runs along the vertical axis; the scene turns the picture so that the coast runs across the frame, as in the view from above. The network of the thesis returns the bounding box of the coastline in the frame; the box in accent is that detection, and its four corners, the small squares, are the features of the thesis's Fig. 4.2. The dashed box in the middle is the desired one: forty pixels across the coast, the full width of the frame, and its corners, the small crosses, are the desired features. The desired positions slide along the frame at a set rate, so the coast scrolls through the image as the craft flies along it.
The thin line from each corner to its cross is the image error. The dashed line from each corner is where the last solution predicts that corner at the end of its horizon. A gust moves the craft from time to time, and the tracking adds its noise to the corners. The chart is the mean error of the features against time, with a mark on its axis at each solution, in the manner of the thesis's Fig. 4.4 and Fig. 4.13.
The model
The camera moves with the velocity of its last solution, in an open loop, and solves again at an event:
the motion of the features in the image under the velocity of the camera
the desired features slide along the frame, so the craft moves along the coast
the tracking reports the features with its noise, which the slider sets
the event: the measured features depart from the predicted ones by more than the bound
the other event: the horizon of the thesis ends
This is the triggering condition of the thesis, with the noise of the visual tracking as the disturbance that takes the real features away from the predicted ones. In the thesis the solution comes from a nonlinear model predictive controller with a horizon of six steps of 0.1 seconds, under the limits of the camera's field of view and of the velocities of the craft; the scene solves no optimal control problem, and a proportional law on the position and the tilt of the box stands in for it at each solution. The thesis solved its problem 28 per cent less often than a controller that solves at every step, in its first simulation.
The control
The slider holds the noise of the tracking, in pixels of the image. With no noise the events come from the bends of the coast, the gusts and the horizon alone. With more noise the measured features depart from the predicted ones more often, and the solutions come closer together. Auto keeps a small noise, and the thesis's horizon of 0.6 seconds.
Where it comes from
The scene draws the background of the paper Coastline Tracking for UAVs Using Event-Triggered Image-Based Visual Servoing Nonlinear Model Predictive Control.