A practical method for counting arbitrary target objects in arbitrary scenes

Abstract

Counting objects such as estimating the number of cells in a microscopic image or the number of pedestrians in a surveillance video is usually accomplished using a counting by detection approach. However, such approaches require explicit object modeling and object detection, and thus often run into problems in the presence of mutual occlusion between objects in the scene. We extend a supervised learning framework that bypasses the challenges in object detection and instead focuses on estimating an object density whose integral over an image region quickly yields the count of objects. Our extensions make it practical for arbitrary objects and scenes. In particular, we automatically determine the area of interest through the motion flow of the objects in the scene, and compensate for perspective effect when it is sensed to be present. Extensive experiments using cells, crowds, traffic, and birds data sets have shown the robustness of the method for a wide range of humanity benefitting applications including security, transportation, biomedicine, ecology, environment, and urban planning.