Aiming at the problem that pharmacists make mistakes due to fatigue in the process of pill sorting, a pill detection algorithm based on improved EfficientDet is proposed in this paper. Firstly, mosaic data enhancement technology is introduced to improve the complexity of sampling data; Then, the backbone network EfficientNet is improved and optimized, and the feature fusion layer of CBMA module is embedded to improve the extraction ability of key features of pills by enhancing learning features; Finally, a cross level data stream from the lower layer to the upper layer is added to the feature fusion part of BiFPN. By making full use of multi-level data, the efficiency of multi-scale feature fusion at different levels is improved. Experiments show that the improved EfficientDet algorithm has a map value of 99.84% in the test, which is 0.65% higher than the original EfficientDet algorithm. At the same time, it also has higher accuracy and better practical application than the target detection networks with better performance such as YOLOv3, YOLOv4 and YOLOv4-Tiny.