Real-time tracking and counting of grape clusters in the field based on channel pruning with YOLOv5s

Accurate fruit counting helps grape wine industry make better logistics and decisions before harvest, and therefore produce higher quality wine. In view of poor real-time performance of the existing fruit tracking and counting methods, and a lack of effective counting methods for cluster-like fruits due to their huge shape variabilities. In this study, an end-to-end lightweight counting pipeline is developed to automate the processing of video data for real-time tracking and counting of grape clusters in field conditions. First, based on channel pruning algorithm, a more lightweight YOLOv5s cluster detection model is obtained, where number of model parameters, model size and floating-point operations (FLOPs) are reduced by 79 %, 76 %, and 58 %, respectively, and the pruned model size is only 3.4 MB. Secondly, the soft non-maximum suppression is introduced in prediction stage to improve detection performance for clusters with overlapping grapes. Test results show that mAP reaches 82.3 % and average inference time is 6.1 ms per image, which effectively reduces model parameters and complexity while ensuring detection accuracy. Finally, online multiple object tracking of clusters is implemented by integrating the detection results and SORT algorithm, where two counting modes are set by introducing counting lines. Test results on 8 videos indicated that the average counting accuracy of the proposed method reached 84.9 %, correlation coefficient with manual counting reached 0.9905, and speed of video processing reached up to 50.4 frames per second (FPS), meeting field real-time requirements. This study provides a timely technical reference for the development of orchard robots to achieve real-time automated yield estimation and accurate crop management decisions.