Abstract: Grassland animal husbandry is the main form of animal husbandry. Cow manure is the by-product of grassland animal husbandry on the vast grassland areas. Since cow manure can serve as an important source of energy, there is a negative impact on grass growth. The distribution of cow manure is also characterized by scattered and concentrated areas in natural grasslands. However, manual picking is still used to collect cow manure in pastoral areas at present, leading to the high labor intensity and cost. Alternatively, computer vision can be expected to apply to the collection of cow manure in recent years. In this article, an image detection model was proposed for cow manure using improved YOLOv5. Firstly, New CSP-Darknet53 was replaced with EfficientFormerV2, in order to improve the boundary sensitivity for low complexity of the model. Experimental results showed that the computational complexity was significantly reduced, compared with the original model. Secondly, PANET was replaced with BiFPN to enhance the feature fusion capability for the high accuracy of detection. There was an increase in the accuracy of detection. Furthermore, CIoU Loss was replaced with Inner IoU Loss, in order to improve the localization accuracy of bounding box regression. Once the ratio value was greater than 1, the larger auxiliary bounding boxes were generated to accelerate the convergence of the sample. A comparison with different ratio values showed that the best performance was achieved when the ratio value was 1.10. From then on, the Bottleneck was replaced in the C3 module with an improved FasterNet block. Leaky Relu was used instead of Relu as the activation function for the FasterNet block module, resulting in significant improvements in the accuracy, recall, and average precision. Additionally, the number of parameters and floating-point operations decreased significantly. The computational complexity was reduced to accelerate the inference speed. Finally, a picking judgment mechanism was added using the YOLOv5 detection box, in order to evaluate the size of the cow manure block and the density of the cow manure group. The cow manure was classified in a certain area for intelligent conditional picking of cow manure. The accuracy of the improved network model was 92.6%, the recall rate was 87.7%, the average accuracy was 87.4%, the parameter count was 4022847B, and the floating-point operation count was 8.1G. The improved model significantly reduced the parameter and operation for the high detection accuracy, in order to significantly improve the performance of the model. A dataset of cow manure was established in pastoral areas, in order to improve data diversity. The collected data was randomly combined using five operations: flipping, scaling and translation, motion blur, random occlusion, and brightness change, in order to enhance data augmentation. Comparison experiments were conducted on the multiple networks and the improved model on the dataset. The experimental results showed that the accuracy of the improved model was superior to other models with the same level of parameter and computational complexity. The improved YOLOv5 model classified the cow manure, according to its different sizes during the inference stage, and then distinguished it using different colored bounding boxes. The recognition and localization of cow manure were achieved in the pastoral areas. The finding can also provide technical support to the intelligent picking vehicles of cow manure.