A Non-cooperative Bistatic Radar Target Detection Method Based on Deep Learning

‍ ‍Non-cooperative bistatic radar target detection usually encountered the problem of weak target detection caused by the weak energy of the received target signal and the difficulty of phase synchronization between echo pulses. To solve this problem， in this paper， we transformed one-dimensional radar data into two-dimensional image data for object detection. The range-slow time image of the radar target is obtained by arranging the pulse compressed radar echo along slow time dimensions. Then the range-slow time images are employed by the YOLO（you only look once）v5s network for feature learning to achieve intelligent detection of non-cooperative bistatic radar targets. Regarding the non-significance of the target to be detected in distance-slow time images and the serious background interference， this paper improved the YOLOv5s network： firstly， the cross-connection pathway is added to the Neck section to ensure the continuous participation of the original information in the training process； secondly， the attention mechanism SENet module is added for strengthening the network’s attention to the target and peripheral information； finally， the Swin Transformer module is introduced in the backbone network to enhance the discovery and characterization of weak targets. Then on the basis of the above work， we proposes a non-cooperative bistatic radar target detection method based on improved YOLOv5s which uses range-slow time image as input. Through numerous of comparative experiments and SNR sensitivity testing experiments， this method obtained 99.1% detection accuracy and 98.8% recall rate on the simulated radar echo data set， which was 4.5% and 4.2% higher than YOLOv5s network， indicating that the improved method can effectively improve the detection performance of YOLOv5s network； and for the target with a 7.1 dB signal to noise ratio of pulse compressed radar echo， the detection rate reaches 98.5%， and the detection probability of the traditional method is only 79.9%， indicating that the proposed method can effectively improve the accuracy and sensitivity of target detection， and has strong application value.