Comparative study on poultry target tracking algorithms based on a deep regression network

Abstract

Poultry tracking is primarily used for evaluating abnormal behaviour and predicting disease in poultry. Offline video is often used to track and record poultry behaviour. However, poultry are group-housed animals. The difficulty of accurately monitoring large-scale poultry farms lies in the automatic tracking of individual poultry. To this end, this paper demonstrates the use of a deep regression network to track single poultry based on computer vision technology. By referring to the Alexnet network, the broiler chicken area of the previous frame and the search area of the next frame were input into the convolutional layer respectively, and the coordinates of the prediction area were obtained by full-connection layer regression. The method was compared with some existing tracking algorithms. Preliminary tests revealed that when compared with MeanShift Algorithm (MS), Multitask learning Algorithm (MIL), Kernel Correlation Filter (KCF), Adaptive Correlation Filters (ACF) and tracking-learn-detection (TLD), the poultry tracking algorithm named TBroiler tracker proposed in this paper has better performance on the overlap ratio, pixel error and the failure rate. TBroiler achieved a mixed tracking performance evaluation (MTPE) of 0.730. The evaluation scores of other methods were 0.362 (MS), 0.355 (MIL), 0.434 (KCF), 0.051 (ACF), and 0.248 (TLD). In addition, the method can be further optimised to improve the overall success rate of verification.