120 likes | 329 Views
Tracking the Trajectory of an Ballistic Projectile with Cameras and Prediction of Landing position by Multilayer Perceptron. By Kin-chung Wong. Motivation. A badminton-playing robot?. What needs to be done. Look at the shuttlecock Estimate the trajectory Go near the trajectory
E N D
Tracking the Trajectory of an Ballistic Projectile with Cameras and Prediction of Landing position by Multilayer Perceptron By Kin-chung Wong
Motivation • A badminton-playing robot?
What needs to be done • Look at the shuttlecock • Estimate the trajectory • Go near the trajectory • Hit the shuttlecock For now, • Restrict the motion to 2-D (a vertical plane) • Assume the robot simply “catches” the shuttlecock at ground level
Objective • Need an algorithm to convert visual information to trajectory estimation • Uses MLP and polynomial approximation • Test different configurations of camera
The watered-down Model • The trajectory of the shuttlecock (replaced by a “projectile” in the project) derived fromF = mg + bv • That is, force = weight + air resistance • Simulation by evaluating ODE • Uses 4° Runge-Kutta method – similar to Euler’s method but more accurate
Scenario #1:Configuration • One camera located in the middle, pointing upward • One camera located in the far right, pointing left • Launch position, velocity are random • No wind
Scenario #1: Result MLP 3-5-2 3rd order Polynomial V(a, b, c) = [1 a b c ab bc ca a2 b2 c2a2b a2c b2a b2c c2a c2ba3 b3 c3 abc] Least-square fit
Scenario #1: Comment • 3rd order polynomial is more accurate due to simplicity of the simulation model • However, a rare input data in test set #2 can cause polynomial to fail! • A single case (projectile thrown almost upward rather than forward) causes 5000% error, which in turn causes overall error SD to shoot up • MLP does not have this problem
Scenario #2: Configuration • Two cameras near the center • Launch position, velocity unknown • No wind
Scenario #2: Result 4-8-2 MLP 3rd order polynomial 4 input variables Again, 3rd order polynomial is better!
Conclusion • Due to simplicity of the simulation model, 3rd order polynomial performs better than MLP • However, MLP output is more stable than high-order polynomial fitting • Future work: Train an MLP to recognize the safety zone of a given polynomial approximation