Abstract:Target recognition is one of the core technologies of the visual perception module of unmanned vehicle automatic driving. At present, target recognition mainly relies on the backbone network to extract features, and then classify and regress the target. In general, the computing and storage capacity of embedded computing platform for unmanned vehicles is limited. In order to reduce the computing power and storage capacity of the backbone network and improve the computing speed and efficiency of unmanned vehicles, this paper makes a comprehensive comparison and analysis of the backbone network for target classification tasks. Focusing on convolution kernel, receptive field, pooling layer, fully connected layer and activation function, taking cifar10 and cifar100 as experimental data, the selection of backbone network operators and network construction are analyzed and compared from the theoretical analysis and data practice level, and the main ideas and practices of feature extraction backbone network construction are summarized and summarized. The results show that the analysis conclusion has a certain theoretical guidance and reference value for the application of the target classification backbone network in the embedded unmanned vehicle system.