Research on Deep Neural Networks for Real-Time Detection and Tracking of Highly Dynamic Targets
Xue Hanqing, Xiaotian Wang, Xue Kai, Chao Chen, Zhang HanYi, Na Jin · 2022 5th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) · 2022
In recent years, artificial intelligence technology represented by deep learning has made major breakthroughs in the field of target recognition, and the application of intelligent target recognition capabilities to the aerospace field is an important direction for future development. However, artificial intelligence technology brings a huge amount of computation and high energy consumption to the aircraft while bringing the target perception capability. However, the aerospace field faces many challenges such as complex environmental interference conditions, lack of data samples, high dynamics, and strong real-time performance. In order to solve the problems above, we purpose a lightweight network which was realized by the separable convolution, ghost bottleneck and structures such as reverse residual modules with different lengths. We applied above purpose method to remote sensing image data. The results indicate that the accuracy of the algorithm is 4.7% higher than traditional lightweight algorithm. To validate the practicability under engineering background, we deployed purposed algorithm to Nvidia edge computing devices Jetson TX2. Experimental data show that our algorithm can achieve 10FPS. It meets the requirements of real-time and accuracy in aerospace applications and has high application value. The technologies described in this study can be used in radar image recognition, multi-modal image analysis, and aerospace intelligent target detection etc., which will effectively improve the intelligence level in the aerospace field and promote the development of equipment intelligence.