The present invention discloses a method for segmenting pedestrians in roadside images using a variable-scale multi-feature fusion convolutional network. It addresses the challenge of significant changes in pedestrian scale by using two parallel convolutional neural networks to extract the local and global features at different scales, and then fusing them to obtain a variable-scale multi-feature fusion convolutional neural network, and this network is trained using roadside pedestrian images to realize accurate pedestrian segmentation, avoiding issues with boundary fuzziness and missing segments commonly found in single-network methods.