Considering the three-dimensional geometric characteristics of the fruit, leaf and branch objects of citrus are real sphere, slice and thin cylinder, and together with the advantage of depth sensors can collect the depth point cloud of the object. A method to recognize citrus fruits based on depth-sphere transversal was proposed. Firstly, the basic principle and the key parameters of the depth-sphere transversal method for spherical fruits feature extraction were proposed. Secondly, point cloud clustering and regional division method were used to obtain isolated and adhering area, and the feature extraction algorithms of isolated fruits and adhering fruits were put forward to fruit and leaf in isolated areas and fruits or leaves in touching areas, respectively. In addition, in-depth data processing and fruits recognition strategy of a complex environment were obtained. According to the Intel RealSense F200 depth sensor parameters, citrus fruit size, close-range detection range, data preprocessing and the requirements for feature extraction algorithm to determine the parameters of the depth-sphere transversal method were carried out. A large number of indoor tests results indicated that the average success rate was 98.4% by the depth-sphere transversal method in isolated area, and the average success rate was 76% in touching region, while the comprehensive success rate was 63.8% in complex environment. The depth-sphere transversal identification method only used the limited depth data points to ensure the accuracy of the original data and at the same time to reduce the amount of computation and the complexity of fruit feature extraction. This can effectively deal with the problem of fruit and leaf occlusion, and achieve the effective distinction between sticking fruits and leaves. The method had a good adaptability to the citrus fruit, which provided a new idea for robots to recognize and locate fruits in complex environment.