Counting pedestrians in crowded scenes with efficient sparse learning

Abstract

Counting pedestrians in crowded scenes provides powerful cues for several applications such as traffic, safety, and advertising analysis in urban areas. Recent research progress has shown that direct mapping from image statistics (e.g. area or texture histograms of people regions) to the number of pedestrians, also known as counting by regression, is a promise way of robust pedestrian counting. While leveraging arbitrary image features is encouraged in the counting by regression to improve the accuracy, this leads to risk of over-fitting issue. Furthermore, the most image statistics are sensitive to the way of foreground region segmentation. Hence, careful selection process on both segmentation and feature levels is needed. This paper presents an efficient sparse training method via LARS (Least Angle Regression) to achieve the selection process on both levels, which provides the both sparsity of Lasso and Group Lasso. The experimental results using synthetic and pedestrian counting dataset show that our method provides robust performance with reasonable training cost among the state of the art pedestrian counting methods.