(Page still under construction)
Using overhead depth sensors as raw data source, we experimented on different technologies to perform pedestrian detection and tracking, going from off-line to real-time capabilities.
Typically, we look at pedestrian tracking as a two-phase job:
- per-frame pedestrian localization
- tracking
Most of our R&D went into point 1. while we outsourced point 2. to scientifc libraries, specifically
- OpenPTV in case of off-line tracking (thanks Alex Liberzon)
- trackpy in case of real-time tracking.
Although our localization methods are published in papers, our codes are, at least for now, not freely available. We are working on localization following two alternative approaches:
- Hierarchical clustering of the depth cloud
see e.g. A.Corbetta et al. Fluctuations around mean walking behaviours in diluted pedestrian flows. Phys. Rev. E, 95 , pp. 032316, 2017. (publications ) - Convolutional Neural Network-based analysis of the depth cloud
see A.Corbetta et al. Weakly supervised training of deep convolutional neural networks for overhead pedestrian localization in depth fields, AVSS17, 2017. (publications ) or this poster