Unweighted Bipartite Matching For Robust Vehicle Counting

Abstract

Intelligent Transportation System (ITS) plays an essential role in smart cities. Through ITS, local authorities could handle enormous traffic flows with minimal effort and solve traffic-related problems such as traffic congestion or traffic regulation violating behaviours. In this work, we designed a system that has the ability to count vehicles moving in specific directions on the road. Such automated systems also have to deal with the diverse weather and instabilities in captured media, making current tracking algorithms become prone to errors. This problem is even more challenging in Vietnam and other developing countries, where traffic on the road is much more complex with the presence of small vehicles such as bicycles and motorbikes, thus tracking algorithms would be more likely to fail. Our proposed method for Track Joining was built on top of deepSORT, incorporating Taylor Expansion and Unweighted Bipartite Maximum Matching to predict missing movements or identify duplicated vehicle tracks, then attempt to merge them. In HCMC AI City Challenge 20201, our whole system outperforms other approaches by achieving the lowest overall RMSE score: an average of 1.39 fails per video segment on a benchmark dataset.