Vehicle Counting: Survey and Experiments

Abstract

Traffic management needs information about traffic to control the flow of transports. With millions of traffic video cameras acting as sensors around the world, collecting the information about traffic flow in real-time is quite easy, but using that information to process and control the traffic flow is a challenge. For detecting vehicles, old methods like inductive loop detectors (ILD), infrared detectors (IRDs), laser sensors, etc. have problems with high cost, efficiency, difficulty, etc. The methods we use in this paper are detection-based counting, regression-based counting. The authors propose a new method that is the combination of two methods above to achieve better results. We also evaluate the viability of using Deep Learning pre-trained models include Faster R-CNN, SSD, YOLO for detection-based. We experiment on 2018 AI CITY CHALLENGE datasets and Vehicles Nepal datasets. Our results show the effectiveness of the combining method in accuracy compares to using each of the methods separately.