Search

CN-122024280-A - Cattle video inventory system and method for pasture management

CN122024280ACN 122024280 ACN122024280 ACN 122024280ACN-122024280-A

Abstract

The invention provides a cattle video inventory system and method for pasture management, wherein the system comprises an edge end and a cloud end, the edge end collects cattle video streams through a camera, cattle features including appearance, movement and gesture features are extracted, local track fragments with time stamps and spatial information are generated based on an improved ByteTrack algorithm, the cloud end receives the features and the track fragments and then carries out multi-mode feature fusion to construct a complete cattle track covering a pasture, space-time logic and multi-source data cross check is carried out, and finally inventory reports including the number of cattle and the abstract of the movement track are generated. The invention realizes non-contact, automatic and high-accuracy checking of the cattle, and improves the efficiency and reliability of pasture management.

Inventors

  • HE YUHAO
  • WANG CAN
  • WU XUANGOU

Assignees

  • 安徽工业大学

Dates

Publication Date
20260512
Application Date
20260127

Claims (10)

  1. 1. A bovine video inventory system for pasture management, the system comprising a rim and a cloud, comprising: The data acquisition and processing module is used for acquiring video streams of the cattle group based on a preset camera at the edge end and preprocessing the video streams; the feature fusion module is used for extracting characteristics of the cattle on the basis of the preprocessed video stream at the edge end, wherein the characteristics of the cattle comprise appearance characteristics, motion characteristics and gesture characteristics; Uploading the characteristics of the cow and the local track segments to the cloud end, wherein the cloud end carries out characteristic fusion on the characteristics of the cow to obtain multi-mode characteristics; and the inventory module is used for constructing an inventory track covering the pasture based on the multi-mode characteristics and the local track segments when the inventory is carried out on the cattle, carrying out integrity check on the inventory track, and generating an inventory report based on the inventory track after the integrity check is carried out.
  2. 2. The system of claim 1, wherein the extraction process of the appearance features is as follows: Inputting the preprocessed video stream into a YOLOv model added with pixel level division branches to obtain a detection frame, the confidence of the detection frame and a division mask; dividing the region in the division mask into three parts, namely a head region, a trunk region and a leg region according to a preset proportion; and extracting color histogram features, texture features and contour shape features in each region, and splicing to generate appearance features.
  3. 3. The system of claim 2, wherein the extraction process of the gesture features is as follows: acquiring two-dimensional coordinates and confidence scores of each key point on the cow body based on a preset HRNet model and the detection frame; Based on the key points, ridge line features, head and neck posture features, limb proportion features and joint angle features are constructed, and posture features are generated by splicing.
  4. 4. A system according to claim 3, wherein the motion feature extraction process is as follows: Sampling optical flow vectors in the segmentation mask by adopting Farneback optical flow algorithm to obtain an optical flow vector set; And extracting a motion speed characteristic, a motion direction characteristic, a motion consistency characteristic and a gait cycle characteristic based on the optical flow vector set, and splicing to generate a motion characteristic.
  5. 5. The system of claim 4, wherein the local track segments are extracted based on a modified ByteTrack algorithm; The improvement to ByteTrack algorithm is as follows: First, the improved ByteTrack algorithm designs a special two-stage association mechanism; the second stage associates the unmatched detection frame of the first stage with the tracking object in the lost state; The second point, the improved ByteTrack algorithm builds a composite similarity model of the position and the feature, and the calculation formula is as follows: Wherein, the To the detection result With tracking objects Is used for the degree of similarity of (c) to (c), In order to detect the result of the test, To the detection result In (a) detection frame and tracking object Is used for the prediction of the cross-correlation of the bounding boxes, To the detection result Cattle only features and tracking objects Cosine similarity of historical average cow characteristics, And (3) with The position similarity weight coefficient and the characteristic similarity weight coefficient are respectively; Thirdly, the improved ByteTrack algorithm designs an active, lost and terminated three-level state system, and is matched with a lost counter, and when a tracking object is not matched for a long time, the tracking object is automatically marked as terminated and a complete track is output; fourth, the improved ByteTrack algorithm adds a composite confidence assessment model, the formula is as follows: Wherein, the For the confidence score of the local trajectory segment, For the number of video frames contained by the local track, For the minimum trusted track length threshold, For the standard deviation of the bovine features in the local trajectory segments over time, The parameters are normalized for the standard deviation of the features, As a function of the minimum value.
  6. 6. The system of claim 1, wherein during the generation of the local track segments, the edge maintains a list of tracked objects including unique local IDs, historical feature sequences, Historical bounding box sequences, tracking status, and loss counters.
  7. 7. The system of claim 1, wherein the multi-modal feature is obtained by: Respectively calculating the reliability weights of the characteristics of the cattle; For the reliability right carrying out normalization again; and carrying out weighted splicing on the cow features to obtain the multi-modal features.
  8. 8. The system of claim 1, wherein the obtaining of the cow trajectory is as follows: In a preset checking time period, calculating cosine similarity of average multi-mode characteristics of the local track segments of different cameras; and splicing the local tracks with cosine similarity exceeding a preset similarity threshold value, so as to obtain the cow tracks, and distributing a global unique ID for each cow.
  9. 9. The system of claim 1, wherein the integrity check comprises a space-time logical check and a multi-source data cross check; the inventory report is in a list form and comprises a global unique ID of the cow, an inventory time period, the number of the cow and a cow track abstract; the cattle body track abstract comprises a coverage area, total activity duration and key time period positions.
  10. 10. A method of video inventory of cattle for pasture management, implemented by the system of any of claims 1-9, comprising: the method comprises the steps that firstly, an edge end and a cloud end are included, wherein the edge end collects video streams of cattle groups based on a preset camera, and the video streams are preprocessed; Extracting characteristics of the cattle on the basis of the preprocessed video stream by the edge end, wherein the characteristics of the cattle comprise appearance characteristics, motion characteristics and gesture characteristics; Uploading the characteristics of the cow and the local track segments to the cloud end, wherein the cloud end carries out characteristic fusion on the characteristics of the cow to obtain multi-mode characteristics; And thirdly, when checking the cattle, the cloud end builds a cattle track covering the pasture based on the multi-mode features and the local track segments, and performs integrity check on the cattle track, and after the integrity check, generates a checking report based on the cattle track.

Description

Cattle video inventory system and method for pasture management Technical Field The invention relates to the technical field of intersection of computer vision, internet of things and intelligent livestock breeding, in particular to a cattle video inventory system and method for pasture management. Background In the background of large-scale development of animal husbandry, accurate inventory and behavior monitoring of the number of cattle in daily management of pastures are important. The traditional inventory mainly depends on manual inventory or contact type identification technologies based on RFID and the like, the labor intensity is high, the efficiency is low, the effect is easily influenced by subjective factors, high-frequency and large-range accurate inventory is difficult to realize, while the later can realize individual identification, but each cattle needs to be provided with an electronic tag, the deployment and maintenance cost is high, misreading or misreading easily occurs in the scene of dense cattle groups and frequent activities, and continuous activity tracks and behavior analysis information of cattle cannot be provided. In recent years, automatic monitoring technology based on computer vision is gradually applied to the field of livestock, but the existing video analysis method still faces significant challenges in complex pasture environments. For example, factors such as similar appearance, mutual shielding, illumination change, weather interference and the like of cattle tend to cause target tracking loss, track breakage or identity confusion, and it is difficult to stably construct a complete individual track under a long-time sequence across cameras. In addition, most systems only depend on single visual characteristics, lack of effective fusion and reliability assessment of multi-dimensional characteristics of cattle, have defects in inventory accuracy and system robustness, and limit practical application in full-scene and automatic pasture management. Disclosure of Invention In order to solve the technical problems mentioned in the background art at present, the invention provides a cow video inventory system and a cow video inventory method for pasture management. For this purpose, the invention adopts the following technical scheme: a cow video inventory system for pasture management, the system comprising a rim and a cloud, comprising: The data acquisition and processing module is used for acquiring video streams of the cattle group based on a preset camera at the edge end and preprocessing the video streams; the feature fusion module is used for extracting characteristics of the cattle on the basis of the preprocessed video stream at the edge end, wherein the characteristics of the cattle comprise appearance characteristics, motion characteristics and gesture characteristics; Uploading the characteristics of the cow and the local track segments to the cloud end, wherein the cloud end carries out characteristic fusion on the characteristics of the cow to obtain multi-mode characteristics; and the inventory module is used for constructing an inventory track covering the pasture based on the multi-mode characteristics and the local track segments when the inventory is carried out on the cattle, carrying out integrity check on the inventory track, and generating an inventory report based on the inventory track after the integrity check is carried out. Further, the extraction process of the appearance features is as follows: Inputting the preprocessed video stream into a YOLOv model added with pixel level division branches to obtain a detection frame, the confidence of the detection frame and a division mask; dividing the region in the division mask into three parts, namely a head region, a trunk region and a leg region according to a preset proportion; and extracting color histogram features, texture features and contour shape features in each region, and splicing to generate appearance features. Further, the extraction process of the gesture features is as follows: acquiring two-dimensional coordinates and confidence scores of each key point on the cow body based on a preset HRNet model and the detection frame; Based on the key points, ridge line features, head and neck posture features, limb proportion features and joint angle features are constructed, and posture features are generated by splicing. Further, the motion characteristic extraction process is as follows: Sampling optical flow vectors in the segmentation mask by adopting Farneback optical flow algorithm to obtain an optical flow vector set; And extracting a motion speed characteristic, a motion direction characteristic, a motion consistency characteristic and a gait cycle characteristic based on the optical flow vector set, and splicing to generate a motion characteristic. Further, the local track segments are extracted based on a modified ByteTrack algorithm; The improvement to ByteTrack algorithm is as follows