The measurement of the traffic load on the road network is essential for calibrating the traffic lights and designing future infrastructural developments. This work is done by so-called enumerators, i.e. human traffic counters. The system was created in order to ease their work by processing the pictures of cameras placed on street furniture with OpenCV (Open Computer Vision Library), and make statistics of the vehicles found.
Due to the outdoor environment the software has to tolerate varying weather conditions, and diverse junctions. The image is processed using multiple methods which perform well in different traffic situations, but result in dissimilar data structures which have to be fused by a dedicated subsystem.
To provide the expected accuracy even in complex situations, multiple hypotheses are created from the input data, which include the right one with higher probability. In the case of overlapping vehicles, more of them might be viable. In this way vehicles get rarely lost, but false positives also occur among the results, which have to be filtered out in possession of more information. In my paper, I also aimed at filtering the duplicates with the highest accuracy.