Cattle body detection based on YOLOv5-ASFF for precision livestock farming

Abstract

Precision livestock farming is a hot topic in the field of agriculture at present. However, due to the diversity of breeding environments, the current intelligent monitoring of animal information still faces challenges. In this study, a YOLOv5-ASFF object detection model was proposed to detect cattle body parts (e.g. individual, head, legs) in complex scenes. The proposed YOLOv5-ASFF consists of two components: YOLOv5 responsible for extracting multi-scale features from sample images, while ASFF was used to adaptively learn fused spatial weights for each scale feature map and fully acquire the features. In this way, the cattle area detection was realized and the generalization of detection model was improved. To verify the applicability and robustness of YOLOv5-ASFF, a challenging dataset consisting of cattle (cow and beef) with complex environments (e.g. different lighting, occlusion, different depths of field, multiple targets and small targets) was constructed for experimental testing. The proposed method based on YOLOv5-ASFFachieved a precision of 96.2%, a recall of 92%, an F1 score of 94.1%, and an mAP@0.5 of 94.7% on this dataset, which outperformed Faster R-CNN, Cascade R-CNN, SSD, YOLOv3 and YOLOv5s. Experimental results showed that the YOLOv5-ASFF method could fully learn more animal biometric visual features, thereby improving the performance of cattle detection model, especially the detection of key parts. Overall, the proposed deep learning-based cattle detection method is favorable for long-term autonomous cattle monitoring and management in intelligent livestock farming.