New visual features and algorithms for real-time camera-based objects detection and recognition in driving environments

Jointly supervised PhD thesis

CAOR / Mines ParisTech + MSP lab Yeungnam University (Korea)

2011-2014

Télécharger le sujet au format PDF

Context and objectives

General context

Detection of real-world objects of interest is a challenging problem and a tremendous amount of research has been done, especially in the last 15 years, on visual object detection and recognition. This topic has become so large that recent surveys generally focus either on a particular application context, or on a particular object type like face, pedestrian, face, and vehicle. Most current methods for real-time visual object detection and categorization are based either on the cascade of AdaBoost-selected features (in line with the pioneering work of Viola and Jones [ViolaJones2001]), or on the main current alternative, which is the use of some classifier (mostly SVM, i.e. Support Vector Machines) applied to Histogram of Oriented Gradient (HOG) proposed in [DalalTriggs2005]. Therefore, it is necessary to research new ideas for designing visual feature and feature dimensionality reduction methods for realtime visual object detection and categorization.

At MSP lab of Yeungnam University (Korea)

The MSP lab of Yeungnam University in Korea has successfully developed several high-level visionbased Advanced Driving Assistance Systems (ADAS) for SL company (which is a rough Korean smaller equivalent of Valeo), and is currently actively researching in the competitive field of robust real-time visual object detection and categorization algorithms for car-embedded ADAS. For instance, MSP lab has recently invented a new feature type, called “moment-based Haar-like feature”, which generalizes conventional Haar-like feature and can improve the performance of object detection. The new feature has been successfully applied to Blind Spot Detection (BSD), and achieved more than 10% improvement of detection rate compared to conventional one. Now, MSP is going to submit the new feature as patent (entitled “moment based Haar-like feature extraction methods and their fast calculation algorithm”) with co-inventor company SL.

At Mines ParisTech (France)

The Robotics Lab (CAOR) of Mines ParisTech is one of the major applied research actors in France on intelligent perception for ADAS, and has a strong and already long (15 years old) tight collaboration with French automotive industry, in particular Valeo. CAOR has a good know-how not only in visual features extraction, but also on feature selection technique and machine-learning for visual recognition of object categories. Regarding the visual features, CAOR has developed its own original family of features, called “control-points features”, first published in [Abramson2005], and further sophisticated by [Moutarde2008]. Moreover, CAOR has recently begun to explore keypoints-based features for object localization and classification, as exposed in [Bdiri2009]. As for feature selection, CAOR now has some experience, in a different application domain [HanMoutarde2011], in one of the most recent dimensionality-reduction sparsity-based methods called Non-negative Matrix-Factorization (NMF) for mapping of many-features characteristics onto a low-dimension space.

PhD thesis objective

The aim of the proposed PhD thesis is to design and develop new and robust methods for real-time camera-based visual object detection and categorization, using latest advances in image analysis and machine-learning, in particular to recognize various object types (e.g. pedestrians, vehicles, animals, etc…) useful in Advanced Driving Assistance Systems (ADAS). Visual objects detection and categorization is a complex process relying on dedicated visual features, feature-selection, and machinelearning algorithms. Therefore, the proposed thesis will investigate the following three aspects: · Research of new visual features · Research of feature dimensionality reduction method · Adaptation of existing machine learning algorithms such as Adaboost and SVM And the new techniques will be tested and tuned on real-world objects detection problems for ADAS.

Candidate profile

Master in a domain related to image and statistical signal analysis * Engineer degree welcome * Good relation and autonomy * Good at writing and for oral presentations * Correct spoken and written English

Particular skills * Scientific knowledge: statistical signal analysis, image processing. * Technical skills: software development (C/C++ /Java/Python)

References

[Abramson2005] Y. Abramson, B. Steux, H. Ghorayeb, “YEF (Yet Even Faster) Real-Time Object Detection“, 2005 Proceedings of International Workshop on Automatic Learning and Real-Time (ALART’05), Siegen, Germany, page 5, 2005. [Bdiri2009] T. Bdiri, F. Moutarde, N. Bourdis and B. Steux, “adaBoost with 'keypoint presence features' for real-time vehicle visual detection”, proc. of 16th World Congress on Intelligent Transport Systems (ITSwc'2009), Stockholm, Sweden, sept. 2009. [DalalTriggs2005] N. Dalal and B. Triggs, “Histograms of Oriented Gradients for Human Detection”, proceedings of 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2005), held in San Diego, CA, USA, 20-26 June 2005. [HanMoutarde2011] Y. Han and F. Moutarde, “Clustering and modeling of Network-level Traffic States based on Locality Preservative Non-negative Matrix Factorization”, accepted at 8th Intelligent Transport Systems (ITS) European Congress, to be held in Lyon (France), 6-9 june 2011. [Moutarde2008] F. Moutarde, B. Stanciulescu, and A. Breheret, “Real-time visual detection of vehicles and pedestrians with new efficient adaBoost features”, proceedings of 'Workshop on Planning, Perception and Navigation for Intelligent Vehicles (PPNIV)' of “2008 International Conference on Intelligent RObots and Systems (IROS'2008)”, Nice, France, September 26th, 2008. [ViolaJones2001] P. Viola and M. Jones, “Rapid Object Detection using a Boosted Cascade of Simple Features”, IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’2001), Volume 1, page 511, Kauai, Hawai, USA, 2001.

Practical informations

Administrative status

The selected candidate will be registered both as Yeungnam University and Mines ParisTech PhD student. The PhD thesis shall be conducted according to the terms of a specific agreement for international joint doctorate supervision to be written and signed by both institutions.

Laboratory (for French part of joint PhD)

Centre de Robotique (CAOR) / Mines ParisTech 60 boulevard Saint Michel, 75272 Paris Cedex 06 http://www.caor.mines-paristech.fr

PhD advisors (for French part of joint PhD)

Fabien MOUTARDE, Tél: (+33) 1.40.51.92.92, Courriel: Fabien.Moutarde@mines-paristech.fr Fawzi NASHASHIBI (?)

Administrative contact (for French part of joint PhD)

Mme Christine Vignaud, Tél : (+33) 1.40.51.92.55, Courriel : christine.vignaud@mines-paristech.fr

 
caor/positions/2011_thesis_obj-recog-yeungnam/imara.txt · Last modified: 2012/04/27 12:13 by MOUTARDE Fabien
Recent changes · Show pagesource · Login