Design And Implementation Of Weaned Piglet Target Tracking System Based On Deep Learning

Weaned piglets are the focus of large-scale pig breeding industry.With the development of intelligent and refined feeding mode,the breeders need to timely grasp the individual behavior changes and health status of weaned piglets under the condition of group raising.Individual tracking of piglets in group environment is the basis of analyzing individual behavior and monitoring individual health of piglets.Due to the problems of light interference,high similarity of piglets appearance,behavior change and occlusion,it is still a technical difficulty to track piglet targets accurately and ensure the uniqueness of identity in complex scenes based on monitoring video.In the current practice of breeding,the tracking and identification of weaned piglets is mainly determined by long-term observation,which is hard to achieve long-term tracking;Large scale pig farms also use RFID electronic ear tag to identify piglets,which has high economic cost and easy to cause short-term stress of piglets when installing electronic ear tag.And large number of weaned piglets is easy to cause identification interference.In addition,the pig tracking method based on artificial marker features has a lot of work to design the underlying features,and has strict requirements for tracking environment.Due to the lack of universality,it is difficult to be used in large-scale pig breeding production.In view of the shortcomings and difficulties of the above methods,a non-contact technology for automatic tracking of weaned piglets is designed.Combined with the current situation of weaned piglet breeding,deep learning was applied to achieve weaned piglet target tracking.In this paper,the weaned piglet target detection model based on deep convolution neural network and the optimized Deep SORT algorithm were used to design the weaned piglet target tracking model,and researched the software platform of weaned piglet target tracking display system.It can realize non-contact and accurate tracking of weaned piglets in the herd environment,and provide technical support for the follow-up analysis of Weaned Piglets’ individual behavior.The main contents of this paper are as follows:(1)Construction of the target detection model of Weaned Piglets: Based on the YOLOv4 target detection algorithm,the deep convolution neural network was designed and the model parameters were trained.CSPDarknet53 was used as the backbone network to enhance the learning ability of CNN,and SPP was used to increase the receptive field of the network.At the same time,the feature maps of 52×52,26×26 and 13×13 scales were used to fuse the feature information of feature maps of different sizes.The PAN structure was adopted.On the basis of the original top-down feature fusion method,the bottom-up process was added to enhance the learning of different scale features of piglets and improve the accuracy of target recognition.The influence of different iterations on the performance of the model was analyzed.The model with excellent performance was selected to detect the target of weaning piglets.The experimental resulted show that the model detection performance is the best when the data enhanced dataset was trained for 20000 iterations,and the average accuracy map and recall were 99.35% and 99.54%,respectively.Based on Center Net target detection algorithm,the improved coding and decoding full convolution network DLA-34 network was used as feature extraction network.By estimating the key points,the center point of the piglet target was found to determine the position of the target and return to the size,position and other attributes of the target.The resolution characteristic image output of 1/4 original image was used.At the same time,Non-maximum suppression was not used for candidate box selection,which reduced the network computation and effectively improved the detection speed and accuracy.The image data and annotation file were input into the neural network,and the model weight file was obtained after off-line iterative training.In the training process,the verification set was tested every five epochs to get the loss function value.The optimal model weight was obtained by comparing the test loss function value of the verification set in the whole training.The experimental result showed that the model detection performance was the best when 105 epochs were trained iteratively,and the average accuracy of map and recall were99% and 78.6% respectively.(2)Build a target tracking model for weaned piglets based on the optimized Deep SORT algorithm: The research adopted the multi-objective tracking strategy based on detection.The weaned piglet target tracking model was designed by combining the offline trained weaned piglet target detection model with the optimized Deep SORT algorithm.The Reid(re-identification)model of Deep SORT algorithm was optimized.The Market1501 data set was selected to train the model of extracting piglet appearance features in the process of optimization tracking,and the re-recognition link was improved.Kalman filter and Hungarian matching algorithm were used to predict the motion state of the next frame and calculate the matching results between the detection and tracking trajectories,so as to realize the real-time detection and tracking of weaned piglets with piglet motion characteristics and appearance characteristics.The multi-target tracking algorithm evaluation indexes such as IDF1,multi-target tracking accuracy MOTA and MOTP,and Rcll,Prcn,FP,FN,FM,MT,ML,PT,IDS were selected to evaluate the function of the designed tracking model.The experimental results showed that the performance of the weaned piglet target tracking model based on the Center Net-det model trained by DLA-34 network combined with the optimized Deep SORT algorithm was the best.Among them,IDF1,MOTA and MOTP reached 68%,88.2% and 81.3%,respectively.(3)Design the software platform of weaning piglet target tracking display system:The system software platform completed the data management,storage and visualization display of piggery environment information data and weaning piglet target tracking process,through the middleware and database deployed in the cloud server.The development of web platform was realized by multithreading,which improved the robustness and efficiency of the platform,and optimized the web page interface.The observation platform was provided to the management personnel of pig farm to reduce the number of times of entering and leaving the pigsty.