Traffic Flow Parameters Collection under Variable Illumination Based on Data Fusion

Figure 1

Pipeline of the proposed method. For image processing, the CycleGAN is used first to generate fake nighttime images based on annotated daytime images. Then, the features of images are extracted by the DLA backbone. For radar processing, the 3D region of interest is utilized for associating the radar detection points with the corresponding object after pillar expansion. Last, the image features are concatenated with the radar features and regressed to other object attributes.