Abstract
We present a modified independent component analysis (mICA) based on the conditional entropy to discriminate unsorted independent components. We make use of the conditional entropy to select an appropriate subset of the ICA features with superior capability in classification and apply support vector machine (SVM) to recognizing patterns of human and nonhuman. Moreover, we use the models of background images based on Gaussian mixture model (GMM) to handle images with complicated backgrounds. Also, the color-based shadow elimination and head models in ellipse shapes are combined to improve the performance of moving objects extraction and recognition in our system. Our proposed tracking mechanism monitors the movement of humans, animals, or vehicles within a surveillance area and keeps tracking the moving pedestrians by using the color information in HSV domain. Our tracking mechanism uses the Kalman filter to predict locations of moving objects for the conditions in lack of color information of detected objects. Finally, our experimental results show that our proposed approach can perform well for real-time applications in both indoor and outdoor environments.
Introduction
Automated visual surveillance faces a persistent challenge: accurately distinguishing humans from other moving objects like animals, vehicles, or swaying trees, especially in "messy" real-world environments. While Independent Component Analysis (ICA) is a powerful tool for extracting features from images, it has a major drawback compared to methods like Principal Component Analysis (PCA): it doesn't naturally "sort" its components by importance. Researchers are often left with a massive haystack of features and no clear way to find the needles that actually help a classifier tell a person apart from a dog. Furthermore, problems like moving shadows and partial occlusions frequently break standard detection algorithms.
Methods
The proposed system architecture integrates several advanced image processing techniques into a cohesive pipeline designed for real-time efficiency.
Moving Object Extraction and Shadow Elimination
The system uses a Gaussian Mixture Model (GMM) to build a dynamic background. To handle shadows—which often fool sensors into thinking an object is larger than it is—the authors apply a "spectral ratio" test. Since shadows change the brightness of a pixel but rarely its color relationships, they can effectively be "canceled out" to preserve the object's true shape.
Occlusion Handling
Recognizing that human bodies are often blocked but heads usually remain visible, the system uses an ellipse-shaped head model. This allows the system to count and locate individuals even when they are walking in a tight group or are partially obscured by environmental elements.
The Modified ICA (mICA)
This is the core innovation of the research. The authors calculate the conditional entropy of each ICA feature relative to the target class (human vs. non-human). Features with low conditional entropy are highly informative because they significantly reduce uncertainty about the object's identity. These top-ranked features are then used to train a Support Vector Machine (SVM) classifier, ensuring the model remains lean and fast.
Tracking
Once detected, humans are tracked using a combination of Kalman Filters to predict movement and HSV color histograms to maintain the identity of the person based on their clothing color, even through temporary occlusions.
Experimental Results
The study utilized a custom dataset captured from 34 different video sequences covering both indoor and outdoor environments. The training set consisted of 1,843 human images and 840 non-human images, while the testing set included 3,178 human images and 2,847 non-human images.
Key findings include:
- High Accuracy: The mICA + SVM approach reached an accuracy of 93.88% when using 30 features, outperforming standard ICA (89.58%) and PCA-based methods (89.65%).
- Computational Speed: The system is remarkably fast, requiring only 1.13 to 1.41 milliseconds to detect an object, making it highly viable for real-time surveillance.
- Model Efficiency: The entropy-based method required significantly fewer Support Vectors (825) compared to other selection methods like Fisher’s Criterion (1157) to achieve better accuracy, proving that high-quality feature selection reduces model complexity.
Conclusion
By treating feature selection as an information theory problem, the authors developed a robust and efficient human detection system. The use of conditional entropy to filter ICA components ensures that only the most discriminative data is processed, leading to high accuracy and real-time performance. Future work will address limitations such as extreme camouflage, massive shadows, and low-resolution captures of distant subjects.
Summary updated