Towards dense people detection with deep learning and depth images

This paper describes a novel DNN-based system, named PD3net, that detects multiple people from single depth image, in real time. The proposed neural network processes image and outputs likelihood map coordinates, where each detection corresponds to Gaussian-shaped local distribution, centered at person’s head. encodes both the number of detected as well their position which 3D can be computed. DNN includes spatially separated convolutions increase performance, runs real-time with low budget GPUs. We use synthetic data for initially training network, followed by fine tuning small amount data. allows adapting different scenarios without needing large manually labeled datasets. Due that, system presented this has numerous potential applications fields, such capacity control, automatic video-surveillance, or groups behavior analysis, healthcare monitoring assistance elderly ambient assisted living environments. In addition, information does not allow recognizing identity scene, thus enabling while preserving privacy. been experimentally evaluated compared other state-of-the-art approaches, including classical solutions, under wide range experimental conditions. achieved results concluding architecture strategy are effective, generalize work scenes those used during training. also demonstrate our proposal outperforms existing methods accurately detect significant occlusions.