Abstract: Aiming at the problem that automatic picking robots cannot quickly and accurately identify fruit and vegetable targets in the complex environments with different lighting, overlapping shadows and large fields of viewm, an improved YOLOv5 algorithm is proposed for fruit and vegetable detection in complex environments. Firstly, Convolutional Block Attention Module is embedded in the CBL module in the backbone network to improve the extraction capability of target features. Secondly, Complete IOU Non-maximum suppression is introduced to improve the regression accuracy by considering the aspect-edge length real difference. Finally, the original YOLOv5 path aggregation network is replaced with a Bidirectional Feature Pyramid Network. The results of this experiment, using Apple as an example, show that the accuracy of the improved YOLOv5 algorithm is 94.7%, the recall is 87%, the average accuracy is 92.5%, which is 3.5% higher than the AP of the original YOLOv5 algorithm, and the detection time under GPU is 11 ms, achieving fast and accurate recognition of fruits and vegetables under complex situations.