Pedestrian segmentation and detection in multi-scene based on G-UNet

Current semantic segmentation methods can obtain the outline of pedestrians, but when pedestrians block each other, the information such as the number, height and central position of the pedestrians in the figure can not be directly obtained. To solve this problem, we propose the G-UNet model algorithm that, in addition to the semantic segmentation backbone, a Gaussian ellipse density kernel detection branch of the pedestrian area is added, so that the center position, height and width of the pedestrian are detected respectively by the maximum point of the kernel, vertical axis and horizontal axis scale. The uniqueness of the maximum point of density kernel solves the detection problem of pedestrian occlusion. Besides, UNet stiffly concatenates the features of the bottom layer and the top layer in a spatial symmetric way so that 50% fixed errors are directly transmitted to the bottom layer. Then, we propose a trainable Soft Connection to obtain the optimal error distribution propagation method. Finally, to solve the problem that the value of the traditional loss function is proportional to the demarcated area of pedestrian, which makes small-scale pedestrians easy to be undetected, an Objective Enhanced Loss was proposed to improve the ability of detecting small-scale pedestrians of network. In the self-built pedestrian segmentation dataset, the experimental results show that the proposed method is effective and superior to other methods. © 2022, Jilin University Press. All right reserved.