In order to solve the problems of complex navigation scenes of inland waterway vessels, few shape and color characteristics of marine radar images, and the difficulty of annotation, an improved YOLOv8 marine radar image target detection method was proposed. Firstly,  to alleviate the issues of annotation errors and model overfitting, a label smoothing strategy was introduced during the model training phase. Then, combining the unique positional prior information of the radar images, a coordinate-based convolution structure was designed to simultaneously extract the shape, color and positional features of the target. To verify the effectiveness and superiority of the proposed method, comparative experiments were conducted on the collected radar images of the Yangtze River channel  under different weather conditions. Results show that the proposed method achieves an accuracy rate of 91.52% while ensuring real-time object detection, with an average accuracy  improvement of  5.17% compared to the classic YOLOv8, which can provide technical support for improving the modernization and intelligent management level of inland waterway shipping.