POLARIZATION-BASED CAR DETECTION

Wang Fan (1) , Samia Ainouz (2) , Fabrice Meriaudeau (3) , Abdelaziz Bensrhair (1)

1 LITIS - Laboratoire d'Informatique, de Traitement de l'Information et des Systèmes
2 INSA Rouen Normandie - Institut national des sciences appliquées Rouen Normandie
3 Le2i - Laboratoire Electronique, Informatique et Image [UMR6306]

Road scene understanding is a vital task for driving assistance systems. Robust vehicle detection is a precondition for diverse applications particularly for obstacle avoidance and secure navigation. Color images provide limited information about the physical properties of the object. This results in unstable vehicle detection caused mainly from road scene complexity (strong reflexions, noises and radiometric distortions). Instead, polarimetric images, characteristic of the light wave, can robustly describe important physical properties of the object (e.g., the surface geometric structure, material and roughness etc). This modality gives rich physical informations which could be complementary to classical color images features. In order to improve the robustness of the vehicle detection purpose, we propose in this paper a fusion model using polarization information and color image attributes. Our method is based on a feature selection procedure to get the most informative polarization feature and color-based ones. The proposed method, based on the De-formable Part based Models (DPM), has been evaluated on our self-collected database, showing good performances and encouraging results about the use of the polarimetric modality for road scenes analysis.

Keywords: Car detection; polarization; feature selection; DPM; road scenes

Domains: Computer Science [cs]; Computer Vision and Pattern Recognition [cs.CV]

Wang Fan, Samia Ainouz, Fabrice Meriaudeau, Abdelaziz Bensrhair. POLARIZATION-BASED CAR DETECTION. IEEE International Conference on Image Processing, Oct 2018, Athena, Greece. ⟨hal-02114561⟩