Jetson NX-oriented Armored Vehicle Fine-grained Recognition

Abstract

Armored vehicles are the main combat equipment on the land battlefield. After discovering armored vehicles, the first task is to identify their fine-grained types, so as to provide more accurate intelligence information for commanders and provide auxiliary decision support for subsequent actions. At present, the research of armored vehicles mainly focuses on target discovery, but ignores the accurate recognition and edge application conditions after discovery. Therefore, to classify the fine-grained armored vehicles, a lightweight adaptive enhanced recognition model is proposed for Jetson NX edge computing device. The model adopts the lightweight RepVGG as the backbone, adds an adaptive enhancement mapping module to the linear classifier head, and transforms the linear classifier head into a nonlinear classifier head, which improves the adaptability of the model in cross-domain transfer learning. At the same time, facing the carrying capacity and power consumption limitation of the combat platform, the re-parameterization technology and the TensorRT optimizer are used to realize the compression, transformation and speed-up of the model on Jetson NX. On the armored vehicle fine-grained recognition dataset, the proposed method achieves good accuracy and fast computational speed.