Research On Livestock Target Detection And Tracking Algorithm Based On Deep Learning

Livestock detection can automatically obtain the location and category information of livestock from the farm monitoring video,while livestock tracking can continuously track individual livestock on the video.The research on livestock detection and tracking can help to realize automatic monitoring of livestock in farms and reduce the labor cost of farms.At present,the reliability of sensor-based monitoring method is poor because of its vulnerability to collision interference.However,the method based on machine vision is also difficult to achieve continuous and effective detection and tracking due to large changes in livestock size,similar target interference,and changes in motion attitude.In this dissertation,the studies are carried out according to the characteristics of livestock from the direction of machine vision and the main contents are as follows:(1)This dissertation inherits the idea of multi-scale feature graph detection of SSD detection algorithm.At the same time,in order to further improve the detection effect of the algorithm for small-scale livestock,this dissertation conducts top-down fusion of hierarchical features based on U-shaped neural network to enhance the semantics of the underlying feature graph.By comparing the mAP values of the detection results of cattle,sheep and pig livestock,the algorithm in this dissertation can achieve effective detection for livestock of different scales,and the fusion feature operation improves the detection results of small-scale sheep.(2)The tracking algorithm in this dissertation is based on Siamese neural network.Considering the high similarity of individual livestock in the farm,in order to improve the distinguishing ability of the tracking algorithm,a hierarchical Siamese neural network structure is proposed by calculating the correlation between the tracking target and the search area on the convolution characteristics of different levels.The results show that this hierarchical matching mechanism can effectively improve the ability to distinguish similar livestock.(3)In order to cope with the challenge of attitude change during livestock movement,an adaptive template updating mechanism is introduced based on response score map.Based on the previous effective tracking results,the dynamic template of the tracking target is updated online and applied to the tracking task together with the first frame template.This method reduces the risk of outdated first frame template information.Tracking experiments verify the effectiveness of the adaptive template mechanism.