A scalable and privacy preserving approach for counting pedestrians in urban environment

Understanding the flow of pedestrians in a city is of paramount importance for urban planning. In this paper, we propose a new approach to pedestrian counting based on using low-cost single-board computers that perform all the video analysis locally. This approach has several advantages: i) the impact on the server-side is minimal when the number of devices is increased, ii) communication requirements are low, and iii) people privacy is assured. A foreground detection algorithm based on keypoint detectors is here proposed to handle the low and unsteady frame rates expected under low-spec hardware. Given a single frame, the algorithm delivers a mask of blobs of potential interest. Several image descriptors are extracted for estimating the number of people. A prototype based on the Raspberry Pi platform was built and installed in a pedestrian street of a mid-size city running the proposed method. Experiments were performed both on data from the prototype and on a public dataset. Results show counting accuracy comparable to related work, while achieving frame rates of ~5 frames per second when running on the Raspberry Pi. We conclude that the proposed system is able to deliver frame rates compatible with typical people counting applications at a low cost while assuring privacy and scalability.