This project continues research aimed at real-time detection and use of pedestrian

traffic flow information to enhance adaptive traffic signal control in urban areas

where pedestrian traffic is substantial and must be given appropriate attention and

priority. Our recent work with Surtrac [12], a real-time adaptive signal control

system for urban grid networks, has resulted in an extended intersection scheduling

procedure that integrates sensed pedestrians and vehicles into aggregate multi-modal

traffic flows and allocates green time on this integrated basis [17]. In this project we

consider the companion problem of providing the pedestrian sensing capability

necessary for effective use of this extended intersection scheduling procedure.

Although some commercial pedestrian detection and counting capabilities do exist,

they typically require the purchase and installation of additional higher resolution

video camera technology, which can double the cost of detection per intersection.

Our interest is in a solution that does not significantly increase infrastructure cost.

The hypothesis investigated in this work is that lower resolution vehicle detection

camera technology can be used to provide a relaxed form of pedestrian count data

that is sufficient for incorporating pedestrian flow information into real-time

intersection scheduling. Specifically, we study the possibility of extracting an

approximate but usable measure of pedestrian “density” from the video stream of a

commercial traffic camera. Our target functionality is the ability to qualitatively

discriminate between “no”, “few” or “many” waiting pedestrians. Contemporary

traffic camera technologies provide resolution as low as 320 × 240 gray scale images

(see Figure 1), together with the ability to specify and monitor a set of occupancy

zones within the image.

Pedestrian detection and counting is not a hard task for humans, but it is challenging

for computers. The challenges include diverse shapes and occlusion among

pedestrians, a dynamic background, and low video quality. First, pedestrians can

have various appearances because of clothing, accessories, assistive devices, and

change of pose while walking. This high intra-class variation, as well as occlusion,

makes pedestrian detection a hard problem. An alternative to classification based

pedestrian detection is to find foreground pixels in each frame of the video, and

analyze those pixels to estimate the number of pedestrians. However, since the

system is deployed in an outdoor environment, shadows caused by moving objects or

sudden change in illumination can create noise that complicates foreground

detection. Moreover, to get a broader view of the intersection, the camera is installed

at a certain height. Thus, pedestrians are small in the images and have fewer details

for computer vision processing.