For accurate recognition of orange fruit targets, a detection algorithm based on YOLOv4 was applied in this research. The results showed that AP (average precision) of YOLOv4 had reached 98.17%, 2.14% and 2.67% respectively higher than SSD and Faster RCNN while recognition rate of traditional image processing algorithms was merely 54.94%. Additionally, the extent of occlusion was proved to have obvious influences on the accuracy of orange detection. The accuracy on slight occlusion conditions appeared to be higher than that on serious occlusion conditions. Generally, YOLOv4 detection algorithm showed its feasibility and superiority on fruit detection in the complex natural environment.