Abstract
In this study, we proposed multiple local features encoded for recognizing the human actions. The multiple local features were obtained from the simple feature description of human actions in video. The simple features are two kinds of important features, optical flow and edge, to represent the human perception for the video behavior. As the video information descriptors, optical flow and edge, which their computing speeds are very fast and their requirement of memory consumption is very low, can represent respectively the motion information and shape information. Furthermore, key local multi-features are extracted and encoded by GA in order to reduce the computational complexity of the algorithm. After then, the Multi-SVM classifier is applied to discriminate the human actions.
Introduction
Human Action Recognition (HAR) is a cornerstone of computer vision, essential for applications ranging from automated surveillance to human-computer interaction. While humans can instantly identify actions like jumping or waving, machines often struggle with the computational complexity required for real-time performance. Existing methods frequently rely on long video sequences and heavy feature extraction processes. This paper addresses the challenge of finding a robust yet simple feature set that can be processed quickly with low memory consumption.
Methodology
The proposed pipeline integrates classical computer vision descriptors with bio-inspired optimization through four main stages:
Dual-Feature Extraction
The system extracts two "simple" features to represent human behavior:
- Canny Edges: These capture the spatial "shape" of the human silhouette, providing a boundary-based representation of the actor.
- Optical Flow: This captures temporal "motion" by calculating the instantaneous velocities of pixels between frames.
Combinational Encoding
To preserve temporal context, features are encoded from a three-frame window (previous, current, and next). These are mapped onto an "angle encoding disk" that quantizes directions into fixed values, transforming visual data into structured vector sets.
Genetic Algorithm (GA) Selection
To manage the massive amount of data produced by video, a Genetic Algorithm is employed. This search heuristic mimics natural selection to evolve a population of feature sets. Through cycles of selection, crossover, and mutation, the GA identifies the sub-grids in video frames with the highest "discriminatory power," effectively pruning redundant data.
Classification
A Multi-SVM (Support Vector Machine) classifier uses these optimized features to categorize the actions.
Experimental Results
The approach was evaluated on two benchmark datasets:
- WEIZMANN Dataset: 81 sequences across 10 action categories.
- KTH Dataset: 599 sequences across 6 action types.
The proposed method achieved state-of-the-art results:
- Accuracy: 99.3% on WEIZMANN and 96.5% on KTH, outperforming contemporary techniques like those by Wang and Suter or Bregonzio et al.
- Efficiency: The GA-optimized feature set (10,000 features) maintained nearly the same accuracy as the full set (15,625 features), representing a 36% reduction in data volume.
Conclusion
By combining Canny edges for shape and Optical Flow for motion, the model achieves high robustness. The use of a Genetic Algorithm allows for significant computational savings without sacrificing accuracy. Future work will focus on unconstrained environments with cluttered backgrounds and the recognition of more complex, multi-stage human activities.
Summary updated