The objective of this article is to study the problem of pedestrian classification across different light spectrum domains (visible and far-infrared (FIR)) and modalities (intensity, depth and motion). In recent years, there has been a number of approaches for classifying and detecting pedestrians in both FIR and visible images, but the methods are difficult to compare, because either the datasets are not publicly available or they do not offer a comparison between the two domains. Our two primary contributions are the following: (1) we propose a public dataset, named RIFIR , containing both FIR and visible images collected in an urban environment from a moving vehicle during daytime; and (2) we compare the state-of-the-art features in a multi-modality setup: intensity, depth and flow, in far-infrared over visible domains. The experiments show that features families, intensity self-similarity (ISS), local binary patterns (LBP), local gradient patterns (LGP) and histogram of oriented gradients (HOG), computed from FIR and visible domains are highly complementary, but their relative performance varies across different modalities. In our experiments, the FIR domain has proven superior to the visible one for the task of pedestrian classification, but the overall best results are obtained by a multi-domain multi-modality multi-feature fusion.
7 February 2020 in IEEE Intelligent Transportation Systems Magazine
The combination of multi-modal image fusion schemes with deep learning classification methods, and particularly with Convolutional Neural Networks (CNNs) has achieved remarkable performances in the pedestrian detection field. The late fusion scheme has significantly enhanced the performance of the pedestrian recognition task. In this paper, the late fusion scheme connected with CNN learning is deeply investigated for pedestrian recognition based on the Daimler stereo vision dataset. Thus, an independent CNN for each imaging modality (Intensity, Depth, and Optical Flow) is used before the fusion of the CNN?s probabilistic output scores with a Multi-Layer Perceptron which provides the recognition decision. We propose four different learning patterns based on Cross-Modality deep learning of Convolutional Neural Networks: (1) a Particular Cross-Modality Learning; (2) a Separate Cross-Modality Learning; (3) a Correlated Cross-Modality Learning and (4) an Incremental Cross-Modality Learning model. Moreover, we also design a new CNN architecture, called LeNet+, which improves the classification performance not only for each modality classifier, but also for the multi-modality late-fusion scheme. Finally, we propose to learn the LeNet+ model with the incremental cross-modality approach using optimal learning settings, obtained with a K-fold Cross Validation pattern. This method outperforms the state-of-the-art classifier provided with Daimler datasets on both non-occluded and partially-occluded pedestrian tasks.