Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Abstract. Pose estimation is critical in equine gait analysis (EGA). Many existing methods rely on veterinarians' visual assessments or motion capture equipment, which are either subjective or time-consuming. This study investigated the feasibility of using stereo 3D machine vision in conjunction with Convolutional Neural Networks (CNNs) for automated EGA. To quantify stride length (SL) and stance duration (SD), a processing pipeline was developed and evaluated using manual annotations in RGB stereo videos and head length (HL) measurements of 40 horses. A DeepLabCut (DLC) model was trained to detect 16 body landmarks in the 2D frames, and their 3D coordinates were reconstructed using Semi-Global Block Matching (SGBM). A pre-trained Faster R-CNN model was fine-tuned to detect hoof and distinguish stance phase from the swing phase in each limb-movement cycle. Given the impracticality of direct infield SL and SD measurements, individual pipeline modules were assessed instead. DLC's RMSE for body landmark detection was 5.14 pixels. Nostril and poll manual annotations and SGBM were used to compute HLs, yielding an R² of 0.9 and an RMSE of 0.020 m regarding 3D reconstruction accuracy. Combining DLC and SGBM, derived HLs had a RMSE and an R² of 0.02 m and 0.5, respectively, reflecting the potential accuracy of SL estimation. The Faster R-CNN model achieved an AP score of 36.46 and a precision of 0.83 after applying a smoothing filter to the hoof detection results. The proposed method demonstrated promising potential as a high-throughput precision tool for EGA and may assist in detection of musculoskeletal diseases and equine breeding.