|
As one of the key technologies for pig behavior recognition, multi-target tracking is of great significance in the field of automated monitoring of large-scale pig farms. Aiming at adhesion and occlusion problems between pigs, this paper proposes an improved IOU-Tracker pig tracking algorithm suitable for real breeding scenarios, and combined with the YOLOv5s network to achieve real-time pig tracking. In the tracking algorithm, the strategy of preferential matching between the Narrow Rectangle of IOU and the Priority Matching by IOU Value to solve the adhesion problem of pigs; it integrates an algorithm of Location Prediction in several frames to predict the reproduced position of the lost target within a certain number of frames, effectively reducing the change of the target ID due to factors such as occlusion and detection frame jitter. Construct pig tracking datasets under different perspectives and scenarios, a total of 7 pig videos, and use the MOT benchmark as the multi-target tracking evaluation standard, also verify on the public dataset of MOT16 and the self-built pig data set. Experiments show that on the public data set MOT16, the algorithm in this paper has reached 48.57% and 88.67% on the MOTA and MOTP indicators respectively; the IDSW times are only 297, which is 59.3% and 49.3% lower than the IOU-Tracker and DeepSort algorithms respectively; On the self-built pig data set, the MOTA index of the algorithm in this paper reached 97.60%, which is 0.37% and 9.39% higher than the IOU-Tracker and DeepSort algorithms respectively; MOTP reaches 98.38%, which is 0.1% and 2.2% higher than IOU-Tracker and DeepSort algorithms respectively; IDSW times are 40 times and 34 times lower than IOU-Tracker and DeepSort respectively; And the FPS reaches 33, which can meet the real-time tracking requirements. Therefore, the algorithm in this paper has a good effect on pig tracking in pig farms, and provides a technical basis for the automatic monitoring of large-scale farms.
|
