Automatic Pig Counting Based on Video Detection and Tracking Algorithm

In the corridor, pigs are more likely to be covered or crowded, the existing methods of automatic pig counting based on video have large errors. In order to tackle these problems, this paper proposes a method of pig counting based on YOLOv5s detector and improved DeepSORT algorithm of behavior tracking. Firstly, DIOU_NMS replaces the ordinary NMS to optimize the output of YOLOv5s detector, which can improve the stability of detection boxes and reduce the missed detection of multiple pigs in parallel. Secondly, using CA model to replace the CV model in the original DeepSORT algorithm to improve the prediction results of kalman filter for trajectories. Finally, setting up the virtual detection area by taking the cross product of the vectors, counting the number of frames and capturing a key frame to accumulate the downside probability for pigs. And the probability as the condition to achieve pig counting in the corridor. The results show that the accuracy of this paper method is high, which can be applied to the pig counting in the breeding farms and provide technical reference for the realization of smart pig breeding.