Ladder rung footfall analysis
Based on bodypart coordinates (limb endpoint, e.g., toe) predicted by a markerless pose tracking program such as DeepLabCut, automatically detect footfalls during the ladder rung experiment. Mid-footfall, start, end, and duration of each detected footfall, as well as the depth (calculated in pixel as mid-footfall location minus the average location of start and end of the footfall), are extracted using the csv file containing continuous bodypart coordinates.
-
Import the csv output from a DeepLabCut model including the estimated bodypart coordinates (or a spreadsheets containing the same header and column structures as a DLC output file). The file should contain at least the coordinates of the limb endpoints (paw / toe) for which you want to analyze the footfalls / slips.
-
Select the slip prediction algorithm:
Threshold: If there is a known spatial threshold of the ladder (y coordinate location), the threshold value can be set in the config.yaml file, so that footfalls detection will only be performed on bodypart coordinates that fall under the threshold.
Deviation: The basic method, applying a peak detection function to coordinate data with minimal preprocessing.
Baseline: Apply an asymmetric least squares baseline to the data to filter out noise and distortion, before detecting footfalls.
- Select the bodyparts to analyze and choose a method for the analysis.
(Select the bodyparts relevant for the experimental design, e.g., hindlimbs only. Select the method for the analysis. )
- The predicted footfalls are then sorted, so that you can validate the prediction by skimming through the results alongside the video frames corresponding to the footfalls / slips, alter their starting and ending times for duration and depth calculation, or add additional footfalls / slips that the program missed, based on coordinate graphs. Zoom in/out of the displayed video frame or coordinate graphs for a clearer view. You should always mark a mistake as footfall before marking its on- and offset times!
(Validate model prediction using the GUI. Top: a frame from original video. Bottom: model prediction of the y-axis location of a bodypart throughout the video duration. The location of the bodypart that led to the current slip prediction is displayed in the graph. This frame is identified as "footfall / slip" by the threshold method, based on pose estimation from DLC. After confirmation, you can select whether it's a "slip" or "fall".)
- If desired, you can mark each identified mistake as either slip or fall for more refined results.
(Slip / fall pop-up)
- Export the automatic output or validated results to a csv file.