Traffic Density Estimation from Highly

12 In this paper, we address the problem of how to accurately estimate the traffic 13 density of road segments from highly noisy image sources. Conventional traffic 14 density estimation techniques from camera feeds typically rely on high quality 15 images. Surprisingly, a large number of live feeds from traffic cameras in devel16 oping regions are highly noisy due to poor camera quality, poor maintenance, lim17 ited field of view, limited network bandwidth (to upload high quality images),blur, 18 multiple reflections and poor illumination effects. We propose a density estimation 19 algorithm which uses a combination of conventional image processing techniques 20 and semi-supervised learning using pre-labeled data to achieve high accuracy with 21 minimal training. Our algorithm supports two different modes of operation for 22 day-time and night-time and is accurate under both settings. We have tested our 23 algorithm based on several hours of real-time traffic feeds from noisy sources in 24 Nairobi, Kenya and Rio De Janeiro, Brazil. 25