This paper proposes a deep learning-based method to detect multiple people from a single overhead depth image with high precision. Our neural network, called DPDnet, is composed by two fully-convolutional encoder-decoder blocks built with residual layers. The main block takes a depth image as input and generates a pixel-wise confidence map, where each detected person in the image is represented by a Gaussian-like distribution, The refinement block combines the depth image and the output from the main block, to refine the confidence map. Both blocks are simultaneously trained end-to-end using depth images and ground truth head position labels. The paper provides a rigorous experimental comparison with some of the best methods of the state-of-the-art, being exhaustively evaluated in different publicly available datasets. DPDnet proves to outperform all the evaluated methods with statistically significant differences, and with accuracies that exceed 99%. The system was trained on one of the datasets (generated by the authors and available to the scientific community) and evaluated in the others without retraining, proving also to achieve high accuracy with varying datasets and experimental conditions. Additionally, we made a comparison of our proposal with other CNN-based alternatives that have been very recently proposed in the literature, obtaining again very high performance. Finally, the computational complexity of our proposal is shown to be independent of the number of users in the scene and runs in real time using conventional GPUs.
This code runs on Tensorflow 2.0
To train the system it is necessary to download the GOTPD database from the following URL:
https://www.kaggle.com/lehomme/overhead-depth-images-people-detection-gotpd1
The Data have to be stored in GOTPD_DATABASE folder and divided in gaussianas, imagenes and validation folders, where we have the training set output images, the training set input images and the validation and test subsets, dividid in another gaussianas and imagenes subfolders.
Once you adapted the data to the folders system, you'll only have to run Train.py and the system will train and save the trained model in DPDnet.h5 (it only save the weights).
The repository stores a DPDnet trained model, it is only a demo model to provide a demonstration, so it won't report the optimal results of the system. To test the only step needed its to run Test.py, and the system will load the demo sequence to show the results obtained with the proposed system.
In the GOTPD_DATABASE/validacion/imagenes and GOTPD_DATABASE/validacion/gaussianas you will find an example test sequence that will load Test.py to show how the system works and compare prediction and groundtruth.
If you make use of this code and/or its related documentation, you are kindly requested to cite the paper:
David Fuentes-Jimenez, Roberto Martin-Lopez, Cristina Losada-Gutierrez, David Casillas-Perez, Javier Macias-Guarasa, Carlos A. Luna, Daniel Pizarro. DPDnet: A Robust People Detector using Deep Learning with an Overhead Depth Camera, Expert Systems with Applications, 2019, 113168, ISSN 0957-4174 https://doi.org/10.1016/j.eswa.2019.113168. (http://www.sciencedirect.com/science/article/pii/S0957417419308851)