METHOD OF SEGMENTING PEDESTRIANS IN ROADSIDE IMAGE BY USING CONVOLUTIONAL NETWORK FUSING FEATURES AT DIFFERENT SCALES

摘要

A method of segmenting pedestrians in roadside image by using a convolutional network fusing features at different scales. To resolve an issue in which pedestrians in an image captured by a roadside smart terminal appear at markedly different scales, the method employs two parallel convolutional neural networks to extract local features and global features of pedestrians at various scales. The local features and global features extracted by the first network and the local features and global features extracted by the second network are fused on a same-level basis, and then secondary fusion is performed on the fused local and global features to obtain a convolutional neural network implementing fusion of features at multiple scales. The network undergoes training, and a roadside pedestrian image is input thereinto to realize pedestrian segmentation. The above method effectively solves the problems of unclear and incomplete segmentation which occur easily in the majority of existing pedestrian segmentation methods based on a single network structure, thereby enhancing the accuracy and robustness of pedestrian segmentation.