A sheep dynamic counting scheme based on the fusion between an improved-sparrow-search YOLOv5x-ECA model and few-shot deepsort algorithm

Abstract

In order to improve the accuracy of sheep counting and avoid the interference of mutual occlusion caused by different moving speed among sheep, the concept of fusion between the improved YOLOv5x model based on the attention mechanism and DeepSort algorithm is proposed. First, the ECA structure that is channel attention mechanism is used to optimize YOLOv5x model to strengthen the ability of capturing global information. Secondly, the sparrow search algorithm based on elite opposition-based learning strategy is used to optimize the learning rate of the detection model, so as to get the weight information of the optimal group to further improve the recognition rate of sheep. In the experiment, 800 high-resolution sheep images augmented by SRGAN network and data augmentation are used as model datasets, and the best weight information obtained by the YOLOv5x-ECA-SSA* model is used to accurately recognize sheep. According to the DeepSort algorithm, the recognized sheep are tracked, predicted and matched optimally. The experimental results show that the test precision of YOLOv5x*, YOLOv5x-ECA* and YOLOv5x-ECA-SSA* based on the SRGAN and data enhancement to train are respectively 95.74%, 96.50% and 97.10%. The error rate of each model combined with DeepSort algorithm to complete sheep dynamic counting is respectively 13%, 12% and 5%. Among them, the YOLOv5x-ECA-SSA* model has the highest mAP and best effect of sheep counting. The result can provide a new theorical method for realizing intelligent dynamic counting and tracking in the grazing process and provide a new technical application for intelligent animal husbandry.