SDNN-RPAM: a spatio-temporal deep neural network based on relative position attention mechanism for video goose counting

For free-range goose breeders, estimating the number of geese can reasonably achieve precise feeding and improve breeding. Complex backgrounds such as weeds, trees, vegetation shadows on the ground, and dynamic movement of geese make it difficult to accurately calculate the number of geese. To address these issues, a spatio-temporal deep neural network based on relative position attention mechanism (SDNN-RPAM) is proposed. First, a series of goose videos are obtained and segmented into continuous frame sequences to construct the dataset. Then, to enable the model to capture long-range dependencies, a multi-head attention mechanism is added to the spatio-temporal deep neural network. This increases the model's focus on the position of the geese, thus reducing the influence of complex backgrounds. Finally, to alleviate the accuracy error caused by the movement of geese, a relative position encoding module is added to the attention mechanism to improve the sensitivity of the model to the movement of geese. Compared with the current some counting methods, the MAE of SDNN-RPAM is 9.22, which is 18.55%, 7.34%, 12.91%, and 1.49% higher than MCNN, CSRNet, Peopleflow, and DSDD, respectively. The proposed method is practical for statistics on the number of geese in free-range farm environments and provides a new feasible tool for density estimation of geese.