Abstract:
Gait recognition, a non-intrusive biometric method effective at a distance, serves as a practical
alternative to traditional identification techniques like fingerprints or facial recognition,
particularly in forensic and surveillance contexts. This study aims to design and compare 2D
and 3D models for gait recognition with the general objective of analyzing their comparative
performance. Specific objectives include: (1) identifying a robust gait dataset recorded across
various environmental conditions and camera angles, (2) extracting 2D and 3D skeletal
keypoints using OpenPose and MediaPipe, respectively, (3) developing and evaluating the
accuracy of 2D and 3D gait recognition models, and (4) analyzing their performance under
varying environmental conditions. The research addresses the effectiveness of dataset
utilization, comparative model accuracy, and key performance influencers. Results show the
2D model excels in occlusion handling, achieving a 0% false positive rate (FPR) for unknown
instances, while the 3D model better generalizes to unseen sequences at higher thresholds, with
a 74.55% FPR. However, the 2D model struggles with single-frame performance (80% FPR),
and the 3D model exhibits significant bias and poor occlusion handling (100% FPR).
Recommendations include enhancing dataset diversity, incorporating temporal features via
LSTM layers, and improving 3D keypoint robustness through occlusion augmentation. This
study provides a replicable framework, offering insights into model trade-offs and advancing
gait recognition for resource constrained environments, forensic applications, and academic
research.
