Improved SSD Model for Pedestrian Detection in Natural Scene

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