Pedestrian Recognition Using Cross-Modality Learning in Convolutional Neural Networks

The combination of multi-modal image fusion schemes with deep learning classification methods, and particularly Convolutional Neural Networks (CNNs) has achieved remarkable performances in the pedestrian detection field. late scheme significantly enhanced performance recognition task. In this paper, connected CNN is deeply investigated for based on Daimler stereo vision dataset. Thus, an independent each imaging modality (Intensity, Depth, Optical Flow) used before CNN?s probabilistic output scores a Multi-Layer Perceptron which provides decision. We propose four different patterns Cross-Modality Networks: (1) Particular Learning; (2) Separate (3) Correlated Learning (4) Incremental model. Moreover, we also design new architecture, called LeNet+, improves not only classifier, but multi-modality late-fusion scheme. Finally, to learn LeNet+ model incremental cross-modality approach using optimal settings, obtained K-fold Cross Validation pattern. This method outperforms state-of-the-art classifier provided datasets both non-occluded partially-occluded tasks.