Person size estimation in image sequences using foreground run-length distributions

Distortions, such as perspective distortion and partial occlusion, causes objects in different locations in the image of a scene appear to have different sizes. We present a new camera calibration method for estimating the dimension of objects, particularly people, in several locations in the image of a scene. Segmentation methods such as background subtraction combined with frame differencing is used to separate background regions from foreground regions, which correspond to transient objects in the scene such as persons. Horizontal run-length features are computed from the binary images of the foreground regions. A feature distribution of the run-length features is accumulated over a certain learning period and maintained for each location in the image. From the feature distributions, object dimensions in each image location are then estimated and expressed as average object width and estimated height. Even with partial occlusion, such as persons behind and aisle, person size estimation works quite well. This calibration method will benefit methods for object detection, robust object tracking, location-specific image filtering, location-specific morphological filtering and estimation of perspective distortion.