Start Date
1-5-2026 12:00 PM
End Date
1-5-2026 1:00 PM
Description
Gait Recognition pipelines have shown promise in recent years with modern models such as SkeletonGait [1] and GPGait [2] demonstrating decent performance on known benchmarks. These Gait Recognition models utilize resource and computationally expensive pipelines for full end-to-end Gait Recognition. This overhead can prove to be a problem for real-time tasks such as on-device recognition. In this project, we aim to reduce one of the most important phases of the Gait Recognition pipeline which is Skeleton and Silhouette generation. We build off of the YOLOv11 architecture [3] simplifying aspects of the existing model to create a model that can efficiently generate the Skeleton and Silhouette with a smaller parameter count and simplified head. We train the model using the MS-COCO 17 [4] dataset and estimate performance using Mean Average Precision on COCO's validation set. We then demonstrate that our pre-processing model was able to achieve mAP scores of .472, .221, .245 which reveals decent performance on Person Detection, Keypoint Estimation, and Mask Prediction tasks respectively with this model. We also achieved a final parameter count of 17.4 million parameters with an inference speed of 16.25 ms. This serves as a solid baseline for an efficient Gait pre-processing model and for simplifying Gait-recognition pipelines.
STRIDER: Human Re-Identification Using Gait Recognition
Gait Recognition pipelines have shown promise in recent years with modern models such as SkeletonGait [1] and GPGait [2] demonstrating decent performance on known benchmarks. These Gait Recognition models utilize resource and computationally expensive pipelines for full end-to-end Gait Recognition. This overhead can prove to be a problem for real-time tasks such as on-device recognition. In this project, we aim to reduce one of the most important phases of the Gait Recognition pipeline which is Skeleton and Silhouette generation. We build off of the YOLOv11 architecture [3] simplifying aspects of the existing model to create a model that can efficiently generate the Skeleton and Silhouette with a smaller parameter count and simplified head. We train the model using the MS-COCO 17 [4] dataset and estimate performance using Mean Average Precision on COCO's validation set. We then demonstrate that our pre-processing model was able to achieve mAP scores of .472, .221, .245 which reveals decent performance on Person Detection, Keypoint Estimation, and Mask Prediction tasks respectively with this model. We also achieved a final parameter count of 17.4 million parameters with an inference speed of 16.25 ms. This serves as a solid baseline for an efficient Gait pre-processing model and for simplifying Gait-recognition pipelines.