To realize accurate feeding of unmanned aquaculture vessels in freshwater ponds, a river crab counting method with YOLO-Crab + improved DeepSORT is developed. First, to address the problems of blurring and low contrast of underwater river crab images, a river crab detection model YOLO-Crab based on YOLOv8 under the preprocessing of CLAHE is proposed.YOLO-Crab adds the coordinate attention mechanism in the backbone to improve the detection precision, and, at the same time, reduces the model magnitude by SimSPPF pooling and GSConv+Slim Neck design to mitigate the model magnitude. The improved DeepSORT algorithm replaces IOU matching with DIOU matching to solve the problem of river crab ID jumping caused by aquatic grass occlusion. Experiments show that the detection precision and F1 of YOLO-Crab model reach 97.3% and 94%, respectively, and the average precision of counting methods is 81%. At the same time, the model was transplanted to Jeston AGX Orin, and the detection accuracy reached 95%, the detection speed was 60 fps, an increase of 50%, and the counting accuracy was 78%, which can provide a reliable basis for accurate feeding of unmanned aquaculture vessels.