Objective To solve the problems of occlusion and difficult identification of small citrus in the nighttime environment and redlize the all-weather intelligent operation of picking robots.

Method A nighttime citrus identification method combined with active light sources was proposed in this paper. Firstly, the best illuminant color was selected by analyzing nighttime citrus images with different color features under active light sources. Then, a nighttime citrus detection model named BI-YOLOv5s was proposed with bi-directional feature pyramid network (Bi-FPN) for multi-scale cross connection and weighted feature fusion to improve the detection performance of occlusion and small citruses. The coordinate attention (CA) module with attention mechanism was introduced to further strengthen the extraction of target location information. Meanwhile, the C3TR module integrated with a Transformer structure was adopted to reduce the computing amount and better extract global information.

Result The precision, recall and average precision of the citrus detection using the BI-YOLOv5s on test set were 93.4%, 92.2% and 97.1%, respectively, with 3.2, 1.5 and 2.3 percent higher than the YOLOv5s, respectively. Moreover, the identification accuracy of the proposed model with an active light source for nighttime citrus was 95.3%, with 10.4 percent higher than the model in the white light environment.

Conclusion The proposed method in this paper has high accuracy for the identification of occlusion and small target citrus in the nighttime environment, and it can provide technical support for nighttime visual identification of intelligent picking of fruits and vegetables.