Pedestrian Detection using Infrared Images and Histograms of Oriented Gradients F. Suard1 , A. Rakotomamonjy1 , A. Bensrhair1 , A. Broggi2
[email protected] 1
Laboratoire d’Informatique, Traitement de l’Information, Syst`emes. INSA de Rouen, Rouen, France 2 Dipartimento di Ingegneria dell’Informazione, Universit` a di Parma, Parma, Italy Intelligent Vehicle Symposium 2006 Tokyo, 14th June 2006
Introduction
HOG Method
Application
Conclusion
Introduction Machine learning and vision system. Histogram of Oriented Gradient [DT05], Classifier : Support Vector Machines [Vap98]. Application : pedestrian detection with infrared images.
Objectives using HOG method for pedestrian detection, extracting windows from infrared images. F. Suard
2
Introduction
HOG Method
Application
Conclusion
Histogram of Oriented Gradient
Introduced by N. Dalal and B. Triggs [DT05] ⇒ representing an image (128 × 64 pixels) with a vector. Computation of local gradient histograms.
F. Suard
3
Introduction
HOG Method
Application
Conclusion
HOG computation steps Original Image :
F. Suard
4
Introduction
HOG Method
Application
Conclusion
HOG computation steps Gradient Orientation and Norm:
F. Suard
4
Introduction
HOG Method
Application
Conclusion
HOG computation steps Cell Splitting:
F. Suard
4
Introduction
HOG Method
Application
Conclusion
HOG computation steps Histogramm normalization, block 1:
Final descriptor: [0.01 0.5 0 0.8] F. Suard
4
Introduction
HOG Method
Application
Conclusion
HOG computation steps Histogramm normalization, block 2:
Final descriptor: [0.01 0.5 0 0.8 0.2 0 0.9 0] F. Suard
4
Introduction
HOG Method
Application
Conclusion
HOG computation steps Histogramm normalization, block n:
Final descriptor: [0.01 0.5 0 0.8 0.2 0 0.9 0 ... 0.6 0.7 0.1 0] F. Suard
4
Introduction
HOG Method
Application
Conclusion
HOG parameters Parameters cell: number of pixels, block: number of cells, overlap, normalization factor (no, L1, L2), histogram: number of bins, weighted vote(gradient magnitude, vote). Exhaustive test for parameters tuning, Dataset : pedestrians and non-pedestrians manually extracted.
F. Suard
5
Introduction
HOG Method
Application
Conclusion
HOG parameters Optimal set of parameters size of cell : 8 × 8 pixels, size of block : 2 × 2 cells, overlap between blocks : 1 cell, normalization factor for block : L2, number of bins per histogram : 8 weigthed vote for histogram : gradient magnitude.
⇒ Descriptor dimension: 3360
F. Suard
6
Introduction
HOG Method
Application
Conclusion
Linear SVM Classifier Data X ∈ Rn Label y ∈ {−1, 1} Decision function f (x) =
Pm
k=1 αk
· yk · hxk , xi + b
Class of X = sign of f (x)
F. Suard
7
Introduction
HOG Method
Application
Conclusion
Single Frame Classification ROC Curve for single frame classification (test dataset: 4400 examples) when size of learning dataset varies :
Confusion matrix (1000) True P N P 2096 54 Prediction N 71 2079 detection 0.9749 accuracy 0.9709 precision 0.9672
F. Suard
8
Introduction
HOG Method
Application
Conclusion
Single Frame Classification ROC Curve for single frame classification (test dataset: 4400 examples) when size of learning dataset varies :
Confusion matrix (1000) True P N P 2096 54 Prediction N 71 2079 detection 0.9749 accuracy 0.9709 precision 0.9672
For 90 % of good recognition : 1 false-positive for 330 windows examined F. Suard
8
Introduction
HOG Method
Application
Conclusion
Misclassification
Examples of bad classification :
misclassification
misclassification
F. Suard
false-positive
false-positive
9
Introduction
HOG Method
Application
Conclusion
HOG applied to infrared images
Application infrared images, pedestrian detection. ⇒ Windows extraction function Particularity of infrared images Warm area (pedestrian head) appears lighter
F. Suard
10
Introduction
HOG Method
Application
Conclusion
Windows extraction Original Image:
F. Suard
11
Introduction
HOG Method
Application
Conclusion
Windows extraction Warm areas:
F. Suard
11
Introduction
HOG Method
Application
Conclusion
Windows extraction Warm area of the second pedestrian:
F. Suard
11
Introduction
HOG Method
Application
Conclusion
Windows extraction Gradient :
F. Suard
11
Introduction
HOG Method
Application
Conclusion
Windows extraction left and right bounds:
F. Suard
11
Introduction
HOG Method
Application
Conclusion
Windows extraction upper bound:
F. Suard
11
Introduction
HOG Method
Application
Conclusion
Windows extraction lower bounds:
F. Suard
11
Introduction
HOG Method
Application
Conclusion
Windows extraction combination and windows extraction (> 1000):
F. Suard
11
Introduction
HOG Method
Application
Conclusion
Windows extraction Classification:
F. Suard
11
Introduction
HOG Method
Application
Conclusion
Results Windows which prediction are over threshold :
f (x) > 0
f (x) > 0.5 F. Suard
12
Introduction
HOG Method
Application
Conclusion
Results
f (x) > 0
f (x) > 0.5
F. Suard
13
Introduction
HOG Method
Application
Conclusion
Results
f (x) > 0
f (x) > 0.5
F. Suard
14
Introduction
HOG Method
Application
Conclusion
Results
f (x) > 0
f (x) > 0.5
F. Suard
15
Introduction
HOG Method
Application
Conclusion
Results
f (x) > 0
f (x) > 0.5
F. Suard
16
Introduction
HOG Method
Application
Conclusion
Conclusion and perspectives + Good results for single frame classification, Answer for pedestrian size variability, Good generalization for pedestrian pose variability. Parameters, Windows extraction. Perspectives Improve performance, reduce computation time, work with sequences. F. Suard
17
Introduction
HOG Method
Application
Conclusion
References
Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In Cordelia Schmid, Stefano Soatto, and Carlo Tomasi, editors, International Conference on Computer Vision and Pattern Recognition, volume 2, pages 886–893, INRIA Rhone-Alpes, ZIRST-655, av. de l’Europe, Montbonnot-38334, June 2005. A. Broggi A. Fascioli P. Grisleri T. Graf M. Meinecke. Model-based validation approaches and matching techniques for automotive vision based pedestrian detection. In Intl. IEEE Wks. on Object Tracking and Classification in and Beyond the Visible Spectrum, San Diego, USA, page in press, June 2005. V. Vapnik. Statistical Learning Theory. Wiley, 1998.
F. Suard
18