SCOUTER

Intelligent video surveillance system
meant to automatically localize the important events

Project cofounded by the Executive Agency for Higher Education, Research,
Development and Innovation Funding

 

04.06.2014

SCOUTER Results Database

In order to obtain the download credentials, please send an e-mail to catalin.mitrea@uti.eu.com.

DOWNLOAD

Publications in established international conferences and journals

C.A. Mitrea, I. Mironică, B. Ionescu, R. Dogaru, A Pseudo-Random Scan perspective to the motion detection paradigm, 4th International Symposium on Electrical and Electronics Engineering (ISEEE), pp. 1-6, 10.1109/ISEEE.2013.6674315, INSPEC13936724, Galati, 2013.

Abstract

A novel motion detection method is proposed and its properties are investigated. It combines pseudo-random scanning of the image with a simple and efficient feature detector inspired from cellular automata dynamics to detect motion as noisy parts of the scanned image. The proposed method (Pseudo Random Scan combined with Clustering Coefficient - PRS+CC) provides several useful features including fast pattern recognition and can be adapted and implemented to detect motion and extract at some point relevant feature. Experimental results suggest computational efficiency when compared to traditional motion detection approaches.

Download

C.A. Mitrea, I. Mironică, B. Ionescu, R. Dogaru, Multiple instance-based object retrieval in video surveillance: Dataset and evaluation, IEEE International Conference on Intelligent Computer Communication and Processing (ICCP), pp. 171-178, ISBN: 978-1-4799-6568-7, INSPEC 14702961, DOI: 10.1109/ICCP.2014.6936970, Cluj, 2014.

Abstract

In this paper we propose a classification-based automated surveillance system for multiple-instance object retrieval task, and its main purpose, to track of a list of persons in several video sources, using only few training frames. We discuss the perspective of designing appropriate motion detectors, feature extraction and classification techniques that would enable to attain high categorization accuracy, and low percentage of false negatives. Evaluation is carried out on a new proposed dataset, namely Scouter dataset, which contains approximately 36,000 annotated frames. The proposed dataset contains 10 video sources, with variable lighting conditions and different levels of difficulty. The video database raises several challenges such as noise, low quality image or blurring, increasing the difficulty of its analysis. Also, the contribution of this paper is in the experimental part, several valuable interesting findings are reported that motivate further research on automated surveillance algorithms. The combination and calibration of appropriate motion detectors, feature extractors and classifiers allows to obtain high recall performance.

Download

C.A. Mitrea, I. Mironică, B. Ionescu, R. Dogaru, Fast Support Vector Classifier for automated content-based search in video surveillance, Signals, Circuits and Systems (ISSCS), pp. 1-4, ISBN: 978-1-4673-7487-3, DOI:10.1109/ISSCS.2015.7203953, Iasi, 2015.

Abstract

In this article we present and test a specialized classifier, i.e., Fast Support Vector Classifier (FSVC), which is employed for multiple-instance human retrieval in video surveillance. Thanks to its low complexity and high performance in terms of computation and speed, FSVC is adapted to ease the generalization of the feature space using only a limited number of samples in the training process. To validate the performance, FSVC is evaluated on two standard video surveillance datasets. It obtains superior or similar results in terms of F2-Score compared to the close related state-of-the-art Support Vector Machines approaches.

Download

C.A. Mitrea, T. Piatrik, B. Ionescu, M. Neville, Retrieval of Distinctive Regions of Interest from Video Surveillance Footage: A Real Use Case Study, 6th International Conference of Imaging for Crime Prevention and Detection ICDP2015, IET INSPEC, UK, London, 2015.

Abstract

The article addresses the issue of retrieving distinctive regions of interest or patterns (DROP) in video surveillance datasets. DROP may include logos, tattoos, color regions or any other distinctive features that appear recorded on video. These data come in particular with specific difficulties such as low image quality, multiple image perspectives, variable lighting conditions and lack of enough training samples. This task is a real need functionality as the challenges are derived from practice of police forces. We present our preliminary results on tackling such scenario from Scotland Yard, dealing with the constraints of a real world use case. The proposed method is based on two approaches: employment of a dense SIFT-based descriptor (Pyramidal Histogram of Visual Words), and use of image segmentation (Mean-Shift) with feature extraction on each segment computed. Tested on real data we achieve very promising results that we believe will contribute further to the ground development of advanced methods to be applied and tested in real forensics investigations.

Download

C.A. Mitrea, S. Carată, M. Ghenescu, SCOUTER: Content-based multiple-instance object retrieval system, IEEE International Conference on Multimedia & Expo Workshops (ICMEW), INSPEC 15330364, DOI 10.1109/ICMEW.2015.7169821, TURIN, 2015.

Abstract

The article presents the main functionalities and outcomes of the SCOUTER project, which is developed by UTI GRUP in partnership with the “Politehnica” University of Bucharest and co-funded by the Romanian R&D National Framework. SCOUTER is built as an offline processing application module for coping with multiple instance objects retrieval task from video surveillance footage (e.g., searching of a general class, a particular object, retrieval of distinctive regions of interest or patterns). Such functionalities are user-driven as the challenges are derived from the practice of police forensics teams on dealing with analysis of large video datasets. Findings and outputs of the presented project are contributed to the understanding of the constraints and issues that are particular to real-world video surveillance datasets and processing systems and further to alleviate analysis efforts of police investigations.

Download

C.A. Mitrea, S. Carată, B. Ionescu, T. Piatrik and M. Ghenescu, Ensemble-based Learning Using Few Training Samples for Video Surveillance Scenarios, fifth International Conference on Image Processing Theory, Tools and Applications IPTA, Orleans, France, 2015. (Accepted).

Abstract

The article targets the task of content-based multiple-instance people retrieval from video surveillance footage. This task is particularly challenging when applied on such datasets as the available samples to train the decisioning system and formulate the query are insufficient (one image, few frames, or seconds of video recording). To cope with these challenges we investigate three established ensemble-based learning techniques, e.g., boosting, bagging and blending (stacking). Such methods are based on a set of procedures employed to train multiple learning algorithms and combine their outputs, while functioning together as a unified system of decision making. The approach was evaluated on two standard datasets (accounting for 16 people searching scenario on ca. 53000 labeled frames). Performance in terms of F2-Score attained promising results while dealing with our current task.

Download

C.A. Mitrea, I. Mironică, B. Ionescu, People Instance Retrieval from Video Footage: At the edge of randomness, UPB Scientific bulletin, Series C Electrical Engineering and Computer Science, 2015 (Submitted).

Abstract

In this paper we address the task of automated multiple-instance people retrieval from large video surveillance databases. These “real-world” datasets rises particular issues in terms of low image quality, multiple image perspectives, variable lighting conditions and most distinctly, the lack of training samples. A previous proposed classification-based method is adapted for experiments on a new dataset. Also, other state-of-the-art descriptors & decisioners pairs are explored and evaluated in terms of F2 average score. Results denote promising performance while the training frames are reduces consistently to unit, where classifiers are showing difficulties to generalize properly.

Download