Pedestrian Detection using Stereovision and Graph Kernels Suard Fr´ed´eric, Guigue Vincent, Rakotomamonjy Alain and Bensrhair Abdelaziz Laboratoire Perception Syst`emes Information (PSI), FRE CNRS 2654 Institut National des Sciences Appliqu´ees (INSA) de Rouen, 1060 Avenue de l’Universit´e, BP08 F-76801 St Etienne du Rouvray Cedex - France Email:
[email protected]
Kernel method
Introduction
We used the SVM classifier and a graph kernel for our problem :
Problem :
• binary classifier : pedestrian or non-pedestrian,
• pedestrian detection,
Learning Set and Margin 4
• artificial vision and machine learning technique ([3], [6], [1], [5], [4]),
3
• adresses the problem of pose variability,
1
• separates the pedestrian from the background with stereovision,
0
−1
2
1
x2
• represents a shape with a graph obtained from the skeleton,
1
0
−1
• classifier : SVM with a graph kernel.
0
−1
Margin
Main advantages of this method :
−2
−4 −4
−3
−2
−1
1
• scale invariability
0
−3
−1
• shape information contained in the graph
0
1
2
x1
• kernel made with the inner products between graphs, obtained with the method of Kashima [2]. It compares two label sequences generated by two synchronized random walks on the two graphs.
Method description This figure shows the steps of objects extraction from the background and the resulting skeleton and associated graph.
Results
Here are some preliminary results that we obtained using our algorithm. • video acquisition : indoor, good light, up to 3 pedestrians per image, • set of 100 pedestrians and 1600 non-pedestrians, • number of training set elements : 1 to 75 per class, • SVM : cross validation, • choose the best parameters for the classifier. The figures below present the classification rate along the number of examples per class in the training set (left) and an example of ROC curve for 10 elements per class (right). 1 0.9
True Positive Rate
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0
0.2
0.4 0.6 False Positive Rate
0.8
1
Conclusions and perspectives
• preliminary results are encouraging, • high classification rate with few patterns for the learning, • scale invariability, • improvements : – classifier performance, – add other labels to the graph to challenge the problem of rotation, – real world situation, – multi-class classifier to detect other object : cars, bicycles.
Graph Designing The method used to transform a black and white image in a graph is as follows. 1. find the skeleton,thanks to a morphological method.
For more information or contact, see http://asi.insa-rouen.fr/∼fsuard/
2. transform this skeleton into a graph made of vertices and edges. A vertice could be defined as a particular pixel, either the end of a skeleton line, or an intersection between many lines of the skeleton. 3. link the vertices together by searching the path between the vertices along the skeleton. Squelette 7
n3 12
18
8 9
14
2
n1 a1 n12
17 5 7
4 234 6 11 5 3
n5 a3 a5 n14 a6 a4n7 n15
a16 a14
a17
n20
19 20 21
10
20 22 26 22 23 16 24 1825 27 19 28 21 23
12 13 14
a24
n25 n2
a15
a2
n13 a8 n16
a7
15
n10 a10 n17 a9 a21n23 a20 n11 n9 a22 n22 a19 n18 a23 n24
n8
References [1] M. Bertozzi, A. Broggi, R. Chapuis, F. Chausse, A. Fascioli, and A. Tibaldi. Shape-based pedestrian detection and localization. In IEEE Intl. Conf. on Intelligent Transportation Systems 2003, 2003. [2] H. Kashima, K. Tsuda, and A. Inokuchi. Marginalized kernels between labeled graphs. In Proceedings of the Twentieh International Conference on Machine Learning, 2003. [3] C. Papageorgiou and T. Poggio. Trainable pedestrian detection. In Proceedings of the 1999 International Conference on Image Processing, pages 35–39, 1999. [4] A. Shashua, Y. Gdalyahu, and G. Hayon. Pedestrian detection for driving assistance systems: Single-frame classification and system level performance. In Proceedings of IEEE Intelligent Vehicles Symposium, 2004.
a13
9 86 11 10
15 17 13 16 11
n6 n19a12 a18 n21 a11 n4
[5] P. Viola, M. Jones, and D. Snow. Pedetrian using patterns of motions and appearance. In IEEE Int. Conf on Computer Vision, pages 734–741, 2003. [6] L. Zhao and C. Thorpe. Stereo and neural network based pedestrian detection. IEEE Trans. on Intelligent Vehicles Transportation systems, 1(3):148–154, 2000.
