Deconvolution Feature Fusion for traffic signs detection in 5G driven unmanned vehicle

Abstract Real-time and accurate recognition of distant traffic signs in a wide visual range is one of the key technologies in 5G driven unmanned vehicles. Most earlier studies focused on improving accurately of short rang traffic signs detection in manned driving assistance system, while little attention has been paid to distant range traffic signs recognition in the unmanned vehicle. At the same time, it is difficult to detect long-distance traffic signs due to their small size. This paper is oriented to the 5G driven unmanned vehicle scene, proposes a novel framework of Deconvolution Feature Fusion based on the backbone of YOLOv3 (DFF-YOLOv3) to enhance the accuracy of distant traffic signs recognition in a wide visual range for unmanned driving. The proposed framework has combined the deep and shallow feature maps to form a fusion module with a wider range and higher accuracy of distant information in a fixed monocular camera. Specifically, the deep feature map is up-sampled by deconvolution and then merged with the shallow feature map. The convolution module is used to learn the feature, and then through the dimensionality reduction of the convolutional layer to form a deconvolution feature fusion module, which finally replaces the original prediction layer to detect the target. Experimental results are provided to validate the framework, which can improve the recognition accuracy of distant traffic signs without reducing the entire visual range in unmanned vehicles. The result shows that the mean accuracy prediction (mAP) of the proposed DFF-YOLOv3 on distant traffic signs is 74.8%, which is higher than other classic detection algorithms.