Conducting traffic survey is a time consuming thus expensive process done by enumerators. The sophistication of today’s image processing algorithms and the computing capacity allows us to fasten the work of the enumerators by counting the vehicles in simple traffic situations and determining their movement directions.
I participated in the construction of a multi-hypothesis based vehicle tracking system that builds traffic statistics by processing the videos of portable cameras. The program creates multiple hypotheses in ambiguous situations. My task is to post filter the hypotheses and locate the uncertain time intervals, where the counting is leaved to the user. Furthermore, I had to compare the statistics generated by the software to the results created manually by the enumerators, thus determine the accuracy.
The created algorithms radically lowers the rate of false positives by filtering the hypotheses belonging to the same vehicles. They locate those specific parts of the video where the level of accuracy expected by the industry may not be reached with enough probability. The performance in this area is measured by the accuracy, so we created a verification system for continuous monitoring which guarantees the quality of the counting process.