Hochschule Karlsruhe Hochschule Karlsruhe - University of Applied Sciences
Hochschule Karlsruhe Hochschule Karlsruhe - University of Applied Sciences

Intelligent Systems Research Group

3D People Detection

Wide area indoor detection and tracking of people is a vital task for many applications, e.g. people counting, customer behavior analysis, ambient assisted living or smart homes.

In our mulit-sensor setup we use multiple low-cost commodity depth sensors using stereo vision. The depth sensors capture the scene from the top-view and share overlapping field of views.

Joint Probabilistic Approach

We developed a probabilistic framework utilizing a generative scene model to jointly exploit the multi-view image evidence, allowing us to detect people from arbitrary viewpoints without the need of a training data set. In Addition, our method is able to operate on a time series of multi-view depth observations, leading to a more robust estimation of the desired probability distribution and an increase in the detection performance. In order to effectively approximate the joint probability distribution of people present in the scene across space and time we make use of mean-field variational inference.

The following video shows the application of our method on the problem of privacy-preserving proximity detection. The top row shows the discrete ground floor probability map inferred by our method without use of any temporal information.

Externer Inhalt

Um diesen Inhalt zu verwenden (Quelle: www.xyz.de), klicken Sie bitte auf Akzeptieren. Wir möchten Sie darauf hinweisen, dass durch die Annahme dieser IFrames Daten an Dritte übertragen oder Cookies gespeichert werden könnten.

Weitere Informationen finden Sie in unserer Datenschutzerklärung.

Future Work

Density estimation for deep learning: Future work will adress the usage of our approach for the annotation of real images. The annotated data will serve as data base for density estimation and the generation of synthetic multi-view depth data. The latter will be used for training deep neural networks.

Related Data Set & Publications

Wetzel J., Laubenheimer A.
MULTIPLE Multi-View Intensity-Depth Data Set for Top-View Indoor People Detection
2020. MULTIPLE

Wetzel J., Zeitvogel S., Laubenheimer A., Heizmann M.
People Detection in a Depth Sensor Network via Multi-View CNNs trained on Synthetic Data
IEEE International Symposium on Electronics and Telecommunication ISETC'20, November 5 - 6, 2020. [pdf]

Wetzel J., Laubenheimer A., Heizmann M.
Temporal Smoothing for Joint Probabilistic People Detection in a Depth Sensor Network
IEEE International Conference on Multisensor Fusion and Integration (MFI 2020). September 14 - 16, 2020. [pdf]

Wetzel J., Laubenheimer A., Heizmann M.
Joint Probabilistic People Detection in Overlapping Depth Images
IEEE Access, vol. 8, 2020. [pdf]

Wetzel J., Zeitvogel S., Laubenheimer A., Heizmann M.
Towards Global People Detection and Tracking using Multiple Depth Sensors
Proceedings, 13th IEEE International Symposium on Electronics and Telecommunication ISETC'18, Timisoara, 2018. [pdf]

Partners & Related Projects

Our research results around 3D people detection were developed in our project WISO3D in collaboration with our partners Vitracom AG and the KIT Institute of Industrial Information Technology (IIIT).