Abstract:

In order to improve the accuracy of the foggy-image pedestrian and vehicle detection, a novel and practical Foggy-image pedestrian and vehicle detection network (FPVDNet) based on the Faster R-CNN is proposed. First, a foggy-density discriminating module (FDM) is proposed to influence the density of the foggy images. In this way, the prediction from the FDM could determine the subsequent operations for different densities of the fog (No-fog, Light fog, and Dense fog). Then, the squeeze and excitation module (SE Module) is designed to use the attention mechanism to improve the feature extraction capability of the network. Meanwhile, the method of the deformable convolution network is applied to add offsets and learn the offsets from target tasks to enhance the transformation modeling capacity of CNNs. Finally, for lack of the annotated fog image dataset, it is necessary to generate a simulated fog image training dataset through the atmospheric scattering model. The simulated foggy image inherits the annotation of the clear image and increases the information on the fog density. Experiments by the proposed FPVDNet are carried out on the 1, 500 real-fog images and 500 real-clear images, with experimental results showing that, compared with the original Faster R-CNN, the mean average detection accuracies are improved 2%~4% by using the FPVDNet.

Key words: fog image object detection, deep learning, Faster R-CNN