A Visual Guidance Algorithm based on DQN

Abstract:
In order to solve the problem of guidance algorithm failure caused by target confrontation when attacking the target, an image space target recognition and guidance method combining deep learning and reinforcement learning is proposed. Through the convolution neural network with depth residuals, the high-dimensional semantic feature information of the target in the global image is extracted. The target to be guided is identified and tracked in real time. Based on the reinforcement learning network and the image space detection information, the guidance strategy in the global image is reasoned and solved. The optimal guidance mode is obtained. The experiment show that the speed of the proposed algorithm is 17 fps, the average recognition accuracy is 98.59%, and the average guidance rate is 90.89%. It can executive identification and guidance in complex environment accurately and quickly, and overcomes the control problem of small unmanned equipment and low-speed missile after the positioning platform is off-line. The algorithm provides carrier platform with the ability of autonomous target recognition and guidance control in the closed data stream.