Lightweight convolutional neural network for aircraft small target real-time detection in Airport videos in complex scenes

Abstract

Airport aircraft identification has essential application value in conflict early warning, anti-runway foreign body intrusion, remote command, etc. The scene video images have problems such as small aircraft targets and mutual occlusion due to the extended shooting distance. However, the detection model is generally complex in structure, and it is challenging to meet real-time detection in air traffic control. This paper proposes a real-time detection network of scene video aircraft-RPD (Realtime Planes Detection) to solve this problem. We construct the lightweight convolution backbone network RPDNet4 for feature extraction. We design a new core component CBL module(Conv (Convolution), BN (Batch Normalization), RELU (Rectified Linear Units)) to expand the range of receptive fields in the neural network. We design a lightweight channel adjustment module block by adding separable depth convolution to reduce the model’s structural parameters. The loss function of GIou loss improves the convergence speed of network training. the paper designs the four-scale prediction module and the adjacent scale feature fusion technology to fuse the adjacent features of different abstract levels. Furthermore, a feature pyramid structure with low-level to high-level is constructed to improve the accuracy of airport aircraft’s small target detection. The experimental results show that compared with YOLOv3, Faster-RCNN, and SSD models, the detection accuracy of the RPD model improved by 5.4%, 7.1%, and 23.6%; in terms of model parameters, the RPD model was reduced by 40.5%, 33.7%, and 80.2%; In terms of detection speed, YOLOv3 is 8.4 fps while RPD model reaches 13.6 fps which is 61.9% faster than YOLOv3.