Wednesday, 21 Feb 2018
You are here:

Subscribe to our Newsletter

How many eyes has a typical person? (ex: 1)

Twitter Feed

EMail Print


Hajer Fradi (Eurecom) visiting Prof. Thomas Sikora (TUB)
28.01.2013 - 28.02.2013
Internal Visits


There is currently considerable interest in visual surveillance systems for crowd analysis. In particular, crowd density estimation has emerged as a major focus of research. In addition to the security reasons, the crowd density measures could provide valuable information for many applications. The main objectives of this exchange were twofold. Firstly, it was proposed to use a crowd density map in order to adjust the level of privacy protection according to the local needs. In particular, adaptive privacy protection filters were built, in which the privacy level gradually decreases with the crowd density. That is based on the observation that the more people are present around a site, the less perceivable and identifiable is a single individual. At the same time, for safety control, video operators need clear visual information in overcrowded areas, in which potentially dangerous events could occur. It is therefore reasonable in many applications to reduce the privacy level in crowded areas compared to spaces with isolated individuals. Secondly, it was proposed to use a crowd density map to enhance people detection and tracking algorithms in highly crowded scenes where delineating individuals is considered a challenging task because of the spatial overlaps. Therefore, the second objective in the collaborative work with TUB was to explore this new promising research direction which consists of using crowd density to complement person detection and tracking in order to improve results.
The main steps of collaborative work between EURECOM and TUB are:
Crowd density map estimation:
It is typically based on the observation that when more local features are close to each other, higher density is obtained. To achieve that, first, local features are extracted. Then, trajectories of these features are built using local optical flow and a forward-backward rejection step. Finally, crowd density maps are estimated from the feature tracks using a symmetric Gaussian kernel function.
Tested features: Scale-invariant feature transform (SIFT), Features from Accelerated Segment Test (FAST), Good Features to Track.
Optical flow method: the Robust Local Optical Flow algorithm.
Crowd density map for privacy preservation:
The estimated crowd density map is valuable information source for privacy preservation. Using an additional head detection step, the degree of data obfuscation for privacy was adapted according to the crowd level.
Tested privacy protection filters: blurring and pixelisation.

Resulting paper:
Hajer Fradi, Volker Eiselein, Ivo Keller, Jean – Luc Dugelay, Thomas Sikora, Crowd ContextDependent Privacy Protection Filters (submitted to the International Conference on Digital Signal Processing 2013, Special Session: Privacy Preservation by Design in Video Content Analysis:
Framework, Architecture, Algorithms and Impact Assessment).
Crowd density map to enhance person detection and tracking in crowded scenes:
Improving person detection and tracking is our second goal in using crowd density maps. Two strategies of the integration step between detection and crowd inputs are proposed: The first idea is to define an adaptive threshold for the confidence score of detections according to the crowd density map. The second idea is to define an adaptive threshold for the overlap between detections according to the crowd density map in the non-maximum suppression step.
In this work, an auto-correction step was proposed, using a learning function that estimate the size of detected bounding box at certain position in the frame. Then, the detections which do not fit the expected size were filtered out. Detection algorithm: the state-of-the-art detector which is an extended form of Histogram of Oriented Gradients (HOG) using multiple scales and resolutions.
Tracking algorithm: Probability Hypothesis Density (PHD) filter using tracking-by-detection.
Datasets: PETS 2009, Crowd Data Driven dataset and UCF dataset.
Future work:
The proposed approach has been implemented and as a future work the evaluation and testing will take place to conclude the improvements achieved by using crowd density maps. The work has been submitted to the International conference on Advanced Video and Signal-Based Surveillance (AVSS 2013). There is still room for improvement of privacy protection by primarily providing better head detectors. So, in the extended work, the goal is to combine the two proposed use cases of the crowd density measure. On the one hand, the crowd density map is used to enhance the head detection step (the same idea as in AVSS submission). On the other hand, it is used to adapt the protection level of personal privacy with incorporation of the improved head detector.
Thus, the novelty of this work is to provide better detection, for better privacy protection. Also, the extended work aims to add objective evaluations of the proposed adaptive protection filters.