People counting with re-identification using depth cameras

Abstract

Low cost real-time depth cameras offer new sensors for a wide field of applications apart from the gaming world. Other active research scenarios as for example surveillance, can take advantage of the capabilities offered by this kind of sensors that integrate depth and visual information.In this paper, we present a system that operates in a novel application context for these devices, in troublesome scenarios where illumination conditions can suffer sudden changes. We focus on the people counting problem with re-identification and trajectory analysis.Automatic people counting offers different application contexts related to security, safety, energy saving or fraud control. Here we go one step further and give hints to extract useful information using depth cameras. The processing of that information allows us to analyze the individuals behavior, particularly if they go away from the typical trajectory, and the problem of re-identifying people.

A Dataset for Persistent Multi-target Multi-camera Tracking in RGB-D

“…Typical approaches focus on obtaining view-invariant appearance features [11], and/or learning matching models specific to a given pair of camera views to be matched [20]. Relatively less studied directions include enhancing re-identification using soft-biometrics like attributes [18], height [25,1], shape [1,15], or movement style [14]. Such techniques are likely to be increasingly important when addressing realistic longer-duration home/office data where identity should be estimated correctly despite that people are likely to change clothes.…”

Predicting Consumers' Locations in Dynamic Environments via 3D Sensor-Based Tracking

“…Analysis of robustness of people tracking in shops with depth sensors is beyond the scope of this paper, but we did obtain fairly large number of trajectories of shoppers in a clothing/ cosmetics section of a real department store: more than 150000 trajectories during 50 days. Similar to the systems, using depth cameras for tracking people in a laboratory [9] and an office [10], our system employs the top-view approach, and humans are subtracted from the background as blobs. Unlike other works on depth camera-based people tracking, that generated background image by averaging [9] or statistical methods [11], our system employs dynamic background model adaptation to support real world requirements.…”

“…Similar to the systems, using depth cameras for tracking people in a laboratory [9] and an office [10], our system employs the top-view approach, and humans are subtracted from the background as blobs. Unlike other works on depth camera-based people tracking, that generated background image by averaging [9] or statistical methods [11], our system employs dynamic background model adaptation to support real world requirements. To the best of our knowledge, up to date adaptive background models were employed either for video data [12] or for combining data from video and depth cameras [13].…”

Treasure Hunt with Intelligent Luminaires

“…The depth data from top view is acquired either placing the depth sensor on the ceiling [9] or performing an affine transformation to the point cloud [10], [11]. The first step in people segmentation and tracking is to estimate the background, which can be achieved by averaging the depth image as in [8], [9] or using statistical method as in [10] where the used Mixture of Gaussian (MoG). After background estimation the region of interest (here people) can be estimated in many ways.…”

“…After background estimation the region of interest (here people) can be estimated in many ways. Easiest way is background subtraction [9], [11] which leaves the remaining foreground pixels presenting the passing people as "blobs". These "blobs" or clusters have to be tracked over the time either with statistical method as in [9] where the extended Kalman filter was used in four state vector representing X,Y and Z coordinates as well as motion in XY plane or in [10] where clusters were tracked with the mixture of Gaussian (GMM) clusters as the parameters for estimated with Expectation Maximisation (EM).…”