Improved SSD Model for Pedestrian Detection in Natural Scene

The indagation improves the SSD network to improve target detection performance in unmanned driving at night. In detection, goal is identify and locate different types of objects present an image. first-level method pulls categorization information position for use by second-level algorithm using featured MAP created deep network. depth characteristics, on other hand, are processed long-distance convolution downsampling. Given lack geographical information, research alludes concept semantic segmentation proposes a improving identification mixing shallow characteristics from backbone with features obtained through sampling. addition, this enhances generation loss function preselection box taking into account peculiarities pedestrian data. Undertakings experiments two data sets provided INRIA Caltech show that improved model USSD reported paper both efficiency speed retrieval.