Research On Visual Positioning And Individual Identification Methods Of Holstein Cattle Based On Deep Learnin

Traditional object detection methods are constrained by factors such as animal welfare,detection time,detection cost,etc.,and the target detection method based on deep learning has the advantages of low error tolerance,high efficiency and not easy interference from human subjective factors through automatic recognition and classification of image analysis.However,there are still the following three problems in the application of animal target detection: first,there is still room for improvement in detection accuracy,especially small target detection;Second,due to the relatively small number of applied research in the field of UAV detection of animals,the UAV remote sensing image dataset is small and small,while the target detection model of automatic animal recognition has high accuracy requirements,and requires a richer training dataset for model training to improve the detection accuracy.Third,the usual method of individual identification is computationally expensive and not conducive to deployment on UAVs.In order to facilitate deployment in UAVs,YOLOv5s was used as the target detection model.Aiming at the detection accuracy problem,four improvement methods are proposed for optimization: the K-means algorithm is used to improve the multi-scale feature proposal ability of the model and accelerate the model convergence.At the Backdone network layer,the first two layers are replaced by using two RepVGG models and an attention mechanism is introduced to enhance the network’s ability to extract features.GSConv was used instead of the original Conv,and the Bottleneck CSP model was replaced with VoVGSCSP to preserve as many hidden connections as possible to improve model accuracy.For sample data,a combination of data augmentation and data augmentation is used to enrich the experimental dataset.Aiming at the problem of high computational cost of individual identification,the tracking of the target is realized by combining YOLOv5 and KCF algorithm,and the single ROI annotation frame obtained from KCF is converted into a set of spatiotemporal trajectories,each set of spatiotemporal trajectories is rescaled and passed to the Inception V3 network,until the 3rd layer of the pool,visual features are extracted from the input frame and fed to the LSTM recurrent neural network,and then recombined with the image frame,after processing this set of spatiotemporal trajectories,The entire input sequence can obtain the final prediction of the ID through the fully connected layer.According to the experimental data,the improved YOLOv5 has high accuracy,recall and AP value,which meets the actual requirements of real-time monitoring scenarios of cattle.The LRCN network is introduced to identify individual cattle with high accuracy and does not require a lot of computing resources.This non-contact real-time monitoring method can effectively improve the intelligence level of pastures,and at the same time is worth promoting for effectively protecting local endangered species and promoting the healthy development of the ecological environment.