TOP  >  Thesis/Dissertation  >  Pedestrian Counting in a Fixed Camera by Combining Regression with Pass Detection

Pedestrian Counting in a Fixed Camera by Combining Regression with Pass Detection

In commercial facilities and event sites, analysis of the numbers of pedestrians in each passage is valuable for marketing and security. Because it is costly to manually count pedestrians in wide area and long-time, automatic counting is required. Our goal is to count pedestrians in an arbitrary installed fixed camera view, such as an existing surveillance camera.

When we count pedestrians manually, it is supposed to count pedestrians who pass through a gate (Virtual Gate: VG) set virtually in the real 3D world. There are some methods for counting pedestrians across a line (Line of Interest: LOI) in the 2D observed image corresponding to the VG. They estimate the numbers of pedestrians by regression with features extracted from the regions of pedestrians passing through the LOI. Owing to the regression approach, they can estimate the numbers of pedestrians accurately even though they are in a cluster. However, in case where there are multiple routes in a camera view, wrong counts are caused by pedestrians that do not pass through a VG because the body parts of them may pass the corresponding LOI.

We propose the pedestrian counting method that combines estimation of the numbers of pedestrians via regression with pass detection via blob tracking. Filtering pedestrians who actually passing a VG by pass detection, the method can enhance the counting precision even in the problematic scenes.

We experimented using videos of fixed cameras in an actual commercial facility, and confirmed the effectiveness of our method. Furthermore, we discuss that the counting results in multiple places can be applied to pedestrian flow estimation.