Traffic jams and economic development are not things that are normally associated with one another. However, if we think of traffic flows and density as measures of infrastructure, we can begin to see the relationship between the two (which CTED Principal Investigator Lakshmi Subramanian discusses in further detail here). Vipin Jain, a CTED student, is researching ways to estimate traffic density in order to effectively manage congestion and to forecast traffic flows.
Traffic Density Estimation from Noisy Sources
Traffic density estimation is an important characteristic required towards congestion detection, traffic management and traffic forecasting. Quantifying traffic density helps in preventing traffic congestion on a real time scale. Moreover, it also helps towards future traffic density forecasts impacting highway capacity estimations. Various approaches like camera/video feeds, car sensor data, loop detector data, and vehicle re identification mechanisms are utilized for density estimation. We base our work on available camera/video feeds which are abundant in today’s highway deployments considering the cheap cost and wide availability aspect.
Camera/video feeds suffer from the below mentioned factors which complicates traffic density estimation.
- Low resolution and poor quality noisy images
- Limited field of view
- Light illumination attributed from multiple reflecting sources
- Density variation across road length
Vehicle identification differs in day and night approaches, where daytime vehicles can be recognized from pre collected sample vehicle images and vehicle matching across the traffic data. Night time traffic estimation is cumbersome due to additional factors like poor lighting and surrounding illumination. Poor lightning involves the camera’s limited quality in night time imagery. The illumination is produced from various factors like vehicle’s reflecting surfaces, billboard reflections, overhead signage and tunnel lightning etc. The illumination becomes denser as one observes a far off point in the field of view as compared to a near point.
We proposed a density estimation algorithm based on the following key steps:
- Calibration and Pre-processing – Manual selection of a road segment under inspection, to be done only once for any road section.
- Pixel Mapping and Graded Measure – Remove extraneous noise to improve evaluation.
- Day time detection – Identification of gray area in an image to correctly identify the free area.
- Night time detection – Identification of white light from vehicles in a weighted manner.
- Graded measure transformation – A Machine learning algorithm to convert the graded pixel scale to actual vehicle number identification.
Our image sources include highly congested road sections in Nairobi, Kenya and Rio De Janeiro, Brazil where we gathered hours of traffic data in the form of image sources and collocated traffic video among several cameras.
The evaluation results show our ability to successfully identify the traffic density in highly noisy images. This work can be leveraged in a very simple & cheap way towards traffic density estimation in developing and or developed countries. Moreover such a system can be used in any traffic management solution towards real-time traffic density estimation, prediction helping to manage the bursty traffic flows.