Ryan Sammon presented his seminar last Tuesday

Ryan Sammon presented his MASc thesis seminar last Tuesday to a small but interested audience.  The title of his talk was “Data Collection, Analysis, and Classification for the Development of a Sailing Performance Evaluation System”.  A photo of Ryan with his thesis supervisor (me!) is below.  Click on the photo for a larger version.  Thanks to Thanos for doing the photographic honours.

 

Advertisements

Ryan Sammon to present next Tuesday

Next Tuesday will the the first summer installment of the seminar series.  Ryan Sammon will be talking about “Data Collection, Analysis, and Classification for the Development of a Sailing Performance Evaluation System”.  The talk abstract follows below.

Time: 1:00pm

Date: Tuesday May 28, 2013

Location: CBY B205 

Data Collection, Analysis, and Classi cation for the Development of a Sailing Performance Evaluation System

Abstract

The performance of teams in vehicle racing scenarios is heavily dependant on the actions of the vehicle crew. In automotive racing, the tracks are xed, alowing the driver to evaluate their performance based upon ideal paths around turns and lap times. In sailing, the courses are not as rigidly de ned, and the ideal course of action for the crew at any time is completely dependant on the environmental conditions at that time. A simple method for evaluating sailing performance is true speed towards the next mark. This method can only be used when sailing in a straight line. Sailing performance depends heavily on how the crew executes turns. Therefore, a system to evaluate the quality of turn execution is desired. The work described in this thesis contributes to the development of such a system. Data was collected using a Blackberry Playbook axed to a J/24 sail boat during races. This data was manually analysed and annotated with three di erent classes of turns: tacks, gybes, and mark roundings. Tests were run using diff erent combinations of pre-processing algorithms and classi cation methods. The pre-processing algorithms include Kalman Fil tering, categorization using quantiles, and residual normalization. The classi cation methods used were nearest neighbour search and various types of multilayer perceptron (MLP) committees. Results indicate that nearest neighbour search with Kalman Filtering alone provides the best classi cation accuracy when part of each turn is presented to the classi cation methods as training data. However, when no data from a turn is presented to the classi cation methods as training data, an averaged probability committee of MLPs with Kalman Fi ltering alone provides the best classi cation accuracy. The results imply collection of additional data would greatly improve classi cation accuracy.