Aiming to solve the problems that strawberry identification and localization were mostly carried out in a simple environment and the identification efficiency was low, the continuous identification and detection of strawberry in a complex environment was studied, and an improved YOLOv3 identification method was proposed. By training a large number of strawberry image data sets, the optimal weight model was obtained. The mean average precision (MAP) of the test set reached 87.51%, among which the average accuracy and the recall rate of mature strawberries was 97.14% and 94.46%, respectively, and that of immature strawberries was 96.51% and 93.61%. In the model test stage, aiming at the problem of strawberry image blurring in the night environment, the recognition accuracy of the original image was significantly improved by using Gamma transform image enhancement. The harmonic mean value (F value) was used as the comprehensive evaluation index, and the actual test results of various identification methods under different fruit numbers, time periods and video tests were compared. The results showed that the improved YOLOv3 algorithm had the highest F value, the average detection time of the picture was 34.99ms, and the average detection frame rate of the video was 58.1f/s, indicating that the recognition accuracy and rate of the model were better than that of other algorithms, and it had good robustness in complex environments such as fruit occlusion, overlap and density. This study can provide theoretical basis for continuous operation of strawberry picking robot under actual working conditions.