A Low-Cost Markerless DeepLabCut-Based Workflow for Spontaneous Gait and Locomotion Analysis in Freely Moving Mice

Summary

We present a low-cost, markerless workflow using DeepLabCut to quantify spontaneous gait and locomotion in freely moving mice. This protocol integrates single-camera video acquisition, markerless tracking of anatomical landmarks, and extraction of spatiotemporal locomotor parameters, providing a non-invasive and accessible framework for motor behavior analysis without specialized equipment.

Abstract

Gait is a widely used functional biomarker for detecting motor alterations across various diseases and conditions, as it reflects changes in coordination, strength, balance, and sensorimotor integration. However, traditional methods to analyze gait in animal models often require expensive equipment, complex setups, or invasive procedures that can alter natural behavior. Here, we present a low-cost, markerless workflow based on DeepLabCut, an open-source pose estimation software, for the quantitative analysis of gait and spontaneous locomotion in freely moving mice. The method relies on single-camera video acquisition, markerless tracking of anatomical landmarks, and extraction of spatiotemporal locomotor parameters, without the need for physical markers or specialized hardware. To demonstrate the protocol's applicability, it was implemented in the triple transgenic mouse model of Alzheimer’s disease (3xTg-AD) as a representative example of application. This approach preserves free movement and minimizes handling-related stress, enabling non-invasive assessment of motor behavior. The protocol is compatible with standard behavioral testing environments. Overall, this method provides an accessible and non-invasive framework for quantitative analysis of gait and locomotion in preclinical research.

Introduction

Gait is a complex motor process that emerges from the real-time interaction of motor, sensory, and cognitive systems, requiring coordinated processing of sensory feedback, postural control, motor planning, and muscle activation to generate stable, adaptive movement¹. Because of this integrative nature, gait alterations are increasingly recognized as functional indicators of neurological dysfunction in both clinical practice and experimental research. In humans and animal models, gait analysis enables detection of changes in mobility, fall risk, frailty, and functional capacity, and has been proposed as a biomarker for disease progression and intervention efficacy across multiple conditions^2,3,4.

In basic research, particularly in murine models, assessing spontaneous gait and locomotor activity represents a widely used strategy to characterize motor phenotypes and to evaluate the impact of physiological perturbations and therapeutic interventions. Objective quantification of locomotor parameters allows the detection of subtle motor alterations that are not reliably captured by visual inspection alone, a limitation that is especially relevant in preclinical studies targeting early or mild phenotypes⁵. In this context, gait analysis has emerged as a valuable tool for investigating alterations in motor behavior and identifying functional changes associated with different experimental conditions, such as neurological and neurodegenerative diseases⁶.

The analysis of motor behavior and gait in murine models has traditionally relied on specialized technological platforms designed to quantify spatial and temporal kinematic variables. Systems such as illuminated walkways, instrumented platforms, and commercial motion-tracking software are commonly used; however, their reliance on costly hardware and dedicated infrastructure significantly limits accessibility for many laboratories^7,8,9,10. Beyond economic considerations, these approaches introduce additional experimental constraints. Several systems require animals to traverse narrow corridors or highly illuminated lanes, imposing forced locomotor patterns that reduce spontaneous behavior, while others depend on physical markers, paint spots, or adhesive tags attached to the limbs. Such interventions can induce stress, alter gait mechanics, and confound the physiological interpretation of motor outcomes^11,12,13,14. Collectively, these limitations highlight the need for methodologies capable of capturing gait under natural, freely moving conditions without reliance on invasive procedures or expensive infrastructure.

Evaluating gait during unconstrained locomotion enables the observation of more natural motor patterns, avoiding external constraints that artificially regulate speed, trajectory, or posture. This approach facilitates the characterization of locomotor behavior under minimally constrained conditions. DeepLabCut is an open-source, deep-learning-based pose estimation framework that uses convolutional neural networks to track user-defined anatomical landmarks without the need for physical markers. By enabling anatomical point tracking from standard video recordings, DeepLabCut allows the extraction of kinematic parameters while substantially reducing cost and experimental complexity^15,16,17,18. Importantly, markerless tracking minimizes experimental interference, making it suitable for studies of spontaneous behavior. This single-camera markerless approach is most suitable for studies requiring accessible, non-invasive quantification of spontaneous locomotor activity and two-dimensional gait-related features under freely moving conditions. However, instrumented gait platforms or three-dimensional motion capture systems are required when direct biomechanical variables, such as ground reaction forces, plantar pressure, joint angles, or full three-dimensional limb kinematics, are needed.

The objective of this article is to present a low-cost, markerless workflow based on DeepLabCut for the non-invasive quantification of gait and spontaneous locomotion in freely moving mice, integrating single-camera video acquisition, anatomical landmark tracking, and extraction of spatiotemporal locomotor parameters within an accessible workflow that does not require specialized hardware or invasive manipulation. This protocol focuses on the methodological implementation and representative outputs of the workflow.

Protocol

All animal procedures were conducted in accordance with the NIH Guide for the Care and Use of Laboratory Animals and were approved by the Bioethics Committee of the Institute of Neurobiology, National Autonomous University of Mexico, under protocol number 117.

1. Study context

1. Use mice under the experimental conditions required for the study. In this protocol article, we use the triple transgenic mouse model of Alzheimer’s disease (3xTg-AD), harboring the APPSwe and tauP301L transgenes on a PS1M146V knock-in background, together with non-transgenic mice as controls, as a representative example of workflow implementation.
2. Genotype all animals before the study begins.
3. Maintain mice with food and water ad libitum under controlled conditions consisting of a 12:12 h light-dark cycle, a temperature of 22 °C ± 2 °C, and relative humidity of 60% ± 5%, under veterinary supervision.
4. Perform recordings during the light phase following a 60 min habituation period, using a single open-field session per animal to preserve spontaneous exploratory behavior and avoid reduced activity associated with repeated exposure to the arena.
   NOTE: In this protocol article, we present representative recordings exclusively to illustrate the workflow and the type of locomotor and gait outputs generated by the method, rather than to support biological comparisons between groups.

2. Materials

1. Use the materials and equipment listed in the Table of Materials and Equipment for the recording and analysis of spontaneous locomotion and gait. See Figure 1 for the overall assembly of the experimental setup.

3. Habituation and recording of spontaneous locomotion

1. Animal preparation and habituation
   1. Carefully transfer the mouse from its housing room to the behavioral testing room at least 60 min before the experiment to allow environmental habituation.
   2. Maintain the behavioral testing room under controlled environmental conditions, including a constant temperature of 20 °C–22 °C, standardized lighting intensity (120–150 lux measured at arena level), and continuous white noise to minimize external auditory stimuli.
   3. During the habituation period, keep the animal in its home cage without introducing additional stimuli or enrichment. Handle animals in accordance with institutional animal welfare guidelines to minimize stress-induced behavioral variability.
   4. Handle the animal using clean gloves and slow, deliberate movements. Lift the animal gently by the base of the tail and support the body with the opposite hand before placing it into the open field arena.
      ​Note: CRITICAL STEP: Inconsistent handling or insufficient habituation time may increase anxiety-related behaviors and introduce variability in locomotor parameters. Ensure identical habituation procedures across all animals and sessions.
2. Open field arena and camera preparation
   1. Before each recording session, thoroughly clean the acrylic open field arena with 70% (v/v) ethanol to remove residual odors and previous scent cues. Allow the surface to dry completely before use.
   2. Handle 70% ethanol in a well-ventilated area, keep it away from heat or ignition sources, and allow the arena to dry completely before use.
   3. Place the open field arena securely on the wooden platform, ensuring that the structure is stable and free of vibrations.
   4. Inspect the transparent acrylic walls and matte floor surface to confirm that they are free of dust, moisture, or reflective residues that could interfere with video tracking.
   5. Position external lateral white light lamps on both sides of the arena at an approximate angle of 30–45° relative to the arena plane to provide uniform illumination across the recording area and minimize shadows, reflections, or excessive brightness. Adjust lamp placement before each session and verify that illumination at the arena level remains within the target range of 120-150 lux.
   6. Position the camera below the arena in a bottom view configuration at a fixed distance of approximately 80 cm from the arena plane, ensuring complete visibility of the recording area. Note that this optimal distance may need to be adjusted depending on the lens aperture and field of view (FOV) of the specific camera to maintain full arena coverage.
   7. Connect the camera to the computer system and initiate the video acquisition software before introducing the animal.
   8. Adjust camera focus and field of view to ensure that the entire arena surface is visible and sharply resolved.
   9. Acquire a reference image for spatial calibration by placing the checkerboard grid or calibration pattern with known dimensions on the arena plane.
      NOTE: CRITICAL STEP: Accurate camera alignment, fixed camera distance, and consistent illumination are essential for reliable markerless pose estimation. Deviations in geometry or lighting conditions can reduce tracking accuracy and compromise data comparability across sessions.
      NOTE: Alternative materials, including stainless steel, may be used for the external support structure, provided that the setup remains stable and recording alignment is preserved.
      NOTE: An opaque arena or alternative wall materials may also be used, particularly when additional control of external visual input is desired, provided that the recording setup remains stable and external visual interference is minimized.
      ​NOTE: The arena dimensions used in this protocol correspond to the configuration employed for recording freely moving mice. These dimensions may be adjusted according to the rodent used and the adopted open-field protocol, provided that complete animal visibility, spatial calibration, and consistent acquisition conditions are maintained.
3. Recording of spontaneous locomotor behavior
   1. Before placing the mouse into the arena, briefly display the animal identifier in the video frame to ensure direct visual association between the recording and the corresponding experimental session.
   2. Initiate video recording before placing the mouse in the center of the open field arena to ensure capture of the initial exploratory behavior from the beginning of the session.
   3. Gently place the mouse in the center of the open field arena.
   4. Record spontaneous locomotion continuously for 3 min without introducing visual, auditory, or tactile disturbances. In the present protocol, only one recording session was performed per mouse.
   5. Monitor the live video feed on the computer screen to confirm uninterrupted recording and proper capture of the animal’s movements.
   6. Periodically verify camera alignment and illumination uniformity between sessions to ensure consistent recording conditions across animals.
   7. At the end of the recording period, stop video acquisition and carefully remove the animal from the arena using the same handling procedure described above.
   8. Return the animal to its home cage and allow it to rest before any further procedures.
      ​NOTE: Avoid repositioning the animal during recording, as external interference may alter spontaneous gait patterns and affect downstream analysis.
4. Cleaning, data storage, and session completion
   1. Clean the open field arena again with 70% (v/v) ethanol after each session and allow it to dry completely before introducing the next animal.
   2. Ensure that no residual ethanol or odor remains on the arena surface before introducing the next animal.
   3. Document the animal identifier, session date and time, experimental condition, and operator name immediately after each recording.
   4. If additional animals are tested, repeat the procedure starting from step 3.1.1 under identical environmental and experimental conditions.
   5. Save each video file in a dedicated folder using a standardized naming convention (e.g., AnimalID_SessionDate_Trial1.mp4).
   6. At the end of the experimental day, shut down the camera system and lighting equipment to prevent drift in illumination settings across sessions.
      NOTE: Although 70% ethanol was used in the present protocol because it is a commonly used disinfectant in this type of procedure, acrylic-compatible disinfectants may also be used to avoid progressive surface deterioration, provided that the arena is completely dry and free of residual odor before the next recording.
      NOTE: When feasible, multiple arenas may be prepared to allow immediate switching between sessions and to facilitate continuous progression of the experiment across animals.
      See Figure 1 for the materials and assembly of the experimental setup, including the wooden platform (Figure 1A), the acrylic arena (Figure 1B), the lateral illumination system (Figure 1C), the bottom-view camera positioning (Figure 1D), and the complete experimental setup (Figure 1E).

Figure 1: Experimental setup for video recording of locomotion in the open field arena. (A) support structure for the arena; (B) acrylic open field arena placed on the platform; (C) lateral illumination to ensure uniform lighting; (D) camera positioning for recording movement from below; (E) complete setup. White arrowheads indicate the orientation of the light source in Figure 1C and of the camera in Figure 1D. Please click here to view a larger version of this figure.

4. Video Processing and Kinematic Analysis

1. Video preparation and quality control
   1. Acquire raw videos in .MOV format and convert them to .avi format for compatibility with DeepLabCut while maintaining constant frame rate and resolution parameters.
   2. Back up all raw video files immediately after acquisition to a dedicated folder on the analysis computer and apply a standardized naming convention (e.g., AnimalID_Session_Date).
   3. Review each video in its entirety to confirm that the recording is complete and free of interruptions, dropped frames, or corruption. Videos with acquisition errors should be excluded from analysis.
   4. Convert the videos to MP4 (H.264) or uncompressed AVI format if the original format is not compatible with DeepLabCut.
   5. Optionally, perform temporal trimming of the video footage to eliminate non-experimental intervals at the beginning and end of the session (e.g., during the introduction or removal of the animal from the experimental arena).
   6. Confirm that the resolution and frame rate match the acquisition parameters used during recording and remain consistent across all sessions.
   7. Verify that the entire arena surface is visible within the video frame and that no body parts are persistently occluded during locomotion.
   8. Place all verified and processed videos into a designated project directory for DeepLabCut analysis.
      ​NOTE: CRITICAL STEP: Inconsistent frame rates, partial arena visibility, or motion blur will directly reduce tracking accuracy. Videos that do not meet quality criteria should not be included in model training or inference.
2. Project creation and configuration
   1. Designate a video library for the project, using either the same videos intended for analysis or a representative subset for network training.
   2. Launch the Python environment and import the DeepLabCut package.
   3. Create a new DeepLabCut project using the command:
   4. deeplabcut.create_new_project('ProjectName','Author','VideoPath').
   5. Define the anatomical landmarks (body parts) of interest in the config.yaml file, ensuring clear and anatomically consistent definitions.
   6. Body part definitions should be optimized for the experimental question and remain identical across all animals and sessions.
3. Frame extraction and body part labeling
   1. Specify start, stop, numframes2extract, and cropping_parameters in the config.yaml file according to video length and spatial layout.
   2. Extract representative frames using: deeplabcut.extract_frames(config_path, 'automatic', 'kmeans'/'uniform').
   3. Review the extracted frames to confirm inclusion of variation in body posture, orientation, limb position, and gait phase coverage across the recording. If specific postures, limb configurations, or locomotor phases are underrepresented, extract additional frames to better represent these conditions.
   4. When specific gait cycle phases are required, manually extract frames using: deeplabcut.extract_frames(config_path, 'manual').
   5. Launch the graphical labeling interface using: deeplabcut.label_frames(config_path).
   6. Label all predefined anatomical landmarks shown in Figure 2, including the representative video frame during acquisition (Figure 2A), the automated body point detection overlaid on the original frame (Figure 2B), and the schematic representation of the defined body landmarks and arena reference points used for analysis (Figure 2C). Follow consistent anatomical criteria (e.g., nose: anterior tip of the snout; tail base: junction between the spine and the tail)
   7. Label all four limbs consistently by placing each landmark at the center of the paw over the plantar pads at ground contact.
   8. Save and inspect the labels using: deeplabcut.check_labels(config_path), and correct inaccurate annotations if needed.
   9. When feasible, ask a second trained evaluator to label a subset of frames to assess inter-rater consistency.
      NOTE: CRITICAL STEP: Inconsistent or anatomically imprecise labeling is a major source of downstream tracking errors. Ensure strict adherence to landmark definitions across all frames.
4. Neural network training
   1. Train the pose estimation model for 200 epochs and obtain a final mean training error (RMSE) of 3.27 pixels and a mean test error of 3.59 pixels, corresponding to stable tracking performance under the acquisition conditions described.
   2. Create the training and evaluation datasets using: deeplabcut.create_training_dataset(config_path).
   3. Adjust the Training Fraction parameter in config.yaml to define the proportion of images used for training.
   4. Verify GPU availability by running: nvidia-smi.
   5. Start network training using: deeplabcut.train_network(config_path, save_epochs=5, epochs=200).
   6. Monitor training loss and stop training once loss values stabilize and no further improvement is observed.
   7. If training is interrupted, resume from the last saved checkpoint.
   8. Evaluate the trained model using: deeplabcut.evaluate_network(config_path) and record the average pixel error.
   9. If the error exceeds the acceptable threshold for the experimental setup, add additional labeled frames or refine annotations and repeat training.
      NOTE: Acceptable pixel error thresholds may vary depending on camera resolution and arena size, but should remain consistent across experiments.
5. Video analysis and gait quantification
   1. Consider points with a likelihood value below 0.3 as missing.
   2. Linearly interpolate missing points only when gaps are ≤3 consecutive frames; exclude longer gaps from analysis.
   3. Define locomotor bouts using a body center velocity threshold of 10 cm/s.
   4. Exclude movements below this threshold and transient events shorter than 3 frames.
   5. Run inference on experimental videos using: deeplabcut.analyze_videos(config_path, [video_path], save_as_csv=True).
   6. Generate labeled videos using: deeplabcut.create_labeled_video(config_path, [video_path]) to visually confirm tracking quality (Figure 2).
   7. Set the pcutoff parameter in config.yaml to 0.6 to apply a confidence threshold during tracking visualization and evaluation.
   8. Import the output CSV files into the statistical computing environment listed in the Table of Materials, together with the corresponding analysis package and the custom scripts provided in Supplementary Material 1 for behavioral quantification.
   9. Convert frame indices to time using the original video sampling rate.
   10. Calibrate spatial dimensions by transforming pixel coordinates into physical units using the known arena dimensions and the tracked arena corner landmarks.
   11. Perform quality control on the tracked coordinates before metric extraction.
   12. Consider points with a likelihood value below 0.3 as missing and replace them with linear interpolation when appropriate.
   13. Identify outlier trajectories during visual inspection and correct or exclude them when necessary.
   14. Define gait segments using synchronized kinematic criteria based on body center displacement and hind limb motion.
   15. Exclude non-locomotor behaviors such as rearing, turning, and grooming.
   16. Manually calibrate and refine the definition thresholds according to the experimental conditions and animal model to retain only robust, continuous locomotor bouts for analysis.
   17. Compute gait metrics derived from geometric and temporal relationships among tracked landmarks using the filtered continuous locomotor segments (Table 1). These included stride length, step length, step width, cadence, stride duration, step duration, stance time, swing time, ipsilateral amplitude, and limb-specific displacement measures.
   18. Extract open field metrics from the full recording using the same tracking data (Table 1). These included total distance traveled, average speed, time spent moving, percentage of time spent moving, number and duration of walking bouts, and cumulative distance covered in the center, periphery, and corner zones.
   19. Compile all outputs into a structured master file for downstream statistical analysis.
   20. Generate visual representations (e.g., trajectories, heatmaps, time-based displacement, gait patterns, and inter-limb coordination plots) using the computational environments listed in the Table of Materials and the plotting scripts described above.
   21. Representative synchronized visualizations of point confidence, gait bout detection, exploratory zone classification, and inter-limb coordination are provided in the Supplementary Video 1 and Supplementary Video 2 to facilitate interpretation of the analytical workflow.
6. Quality control and data backup
   1. Visually inspect a random subset of labeled videos to confirm anatomical consistency across frames and sessions.
   2. Compare predicted landmark positions with annotated labels to verify model stability and generalization.
   3. Save a complete copy of the project, including config.yaml, original labels, trained network weights, and output files.
   4. Document the software versions used in the workflow, including DeepLabCut, Python, PyTorch, CUDA, and the statistical computing environment, together with GPU specifications and the analysis date.
   5. Generate a summary report of average errors and processing times from evaluation-results.csv.
   6. Back up the entire project to a local archive and to an institutional server or secure cloud storage.
7. Computational environment and reproducibility resources
   1. Capture original video recordings in .MOV format at a resolution of 1080 × 1920 pixels and approximately 60 FPS.
   2. Preprocess and transcode the videos to .avi format.
   3. During this conversion step, crop the region of interest, resize it to a square resolution of 720 × 720 pixels, and standardize the frame rate to a constant 60 FPS to ensure accurate temporal quantification during subsequent kinematic analysis.
   4. Implement the markerless tracking workflow using DeepLabCut version 3.0.0rc13 in a Python environment configured with Python 3.12.12, PyTorch 2.5.1, and CUDA 11.8.
   5. Train the pose estimation model for 200 epochs and obtain a final mean training error (RMSE) of 3.27 pixels and a mean test error of 3.59 pixels.
   6. Perform post-tracking analysis using the statistical computing environment and the custom scripts provided in Supplementary Material 1 for gait and open-field quantification.
   7. Use a body center velocity threshold of 10 cm/s to identify walking bouts. Fill gaps of up to 3 frames between locomotor epochs and exclude transient movements shorter than 3 frames to retain continuous gait segments for analysis.
   8. Provide the relevant config.yaml settings, representative analysis scripts, example video and tracking files, and model weights in Supplementary Material 1.
   9. Provide additional implementation details, model evaluation metrics, hardware specifications, and case-specific analysis parameters in Supplementary Material 1, where applicable.
      NOTE: Adjust these parameters according to the acquisition conditions used in each experiment.

Figure 2: Representative workflow for markerless body point tracking and gait analysis from video recordings. (A) Representative video frame of a mouse freely moving within the open field arena during acquisition. (B) Example of automated body-point detection overlaid on the original video frame after pose estimation, illustrating the anatomical landmarks used for analysis. (C) Schematic representation of the defined body landmarks and arena reference points used for gait reconstruction, spatial calibration, and quantitative locomotor analysis. Landmarks and arena reference points: snout, throat, body center, body center left (bcl), body center right (bcr), left hip (hipl), right hip (hipr), tail base, tail center, tail tip, left forepaw (lfp), left hindpaw (lhp), right forepaw (rfp), right hindpaw (rhp), top left (tl), top right (tr), bottom left (bl), and bottom right (br). Please click here to view a larger version of this figure.

Results

Following application of the tracking and analysis workflow, automated processing reconstructed spontaneous open field locomotion using the defined body center landmark and generated quantitative outputs suitable for downstream analysis. Representative trajectory plots illustrate how the protocol captures spatial displacement across the arena, temporal occupancy, and speed dynamics during free exploration. The open field variables shown in this section were calculated from the full recording, and representative recording...

Discussion

This study presents a low-cost, markerless approach for analyzing spontaneous locomotion and gait in freely moving mice using standard video acquisition and DeepLabCut-based pose estimation within an accessible workflow. The workflow allows the extraction of locomotor and gait-related variables from video recordings while preserving natural locomotor behavior and reducing experimental interference. In this sense, the protocol offers a practical and accessible alternative for preclinical studies of motor behavior.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
70% ethanol	Generic	N/A	Arena cleaning between sessions
Adjustable white-light lamps	Generic	N/A	Maintain constant illumination (100–150 lx)
Black curtains/glare control	Generic	N/A	Reduce external light/reflections
Checkerboard grid template	Generic	N/A	Spatial calibration reference image
DeepLabCut (v3.0.0rc13)	Open-source	N/A	Pose estimation software
Digital video camera (HD, 16:9; ≥60 fps)	Generic	N/A	Fixed settings across sessions
Nvidia GPU workstation	Custom build	N/A	Hardware for model training/inference
Lux meter	Generic	N/A	Verify illumination at the arena level
PyTorch	Pytorch Foundation	N/A	Framework required by DeepLabCut
Transparent acrylic open-field arena (42 × 42 × 42 cm)	Custom-made / In-house	N/A	Open-field arena
Under-arena camera mount (90° view)	Custom-made / In-house	N/A	Stable mount for bottom recording
CUDA	Nvidia corporation	N/A	GPU-accelerated parrellel proccesing
R programming language	R Foundation	N/A	Programming language for data analysis