Crowd Noise Classification via Spectral Time Domain Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing surveillance systems face limitations in detecting abnormal events in public places due to environmental conditions and noise interference, with combined audio and video features often reducing performance rather than improving it.
Innovation Solution
A system and method for classifying noise data of human crowds using a combination of spectral and time domain features, such as Mel-Frequency Cepstral Coefficients, Spectral Flatness Measure, Short-Time Energy, and Zero Cross Rating, fused at score levels using Support Vector Machine classification models to enhance detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If audio based surveillance systems are used to detect events in noisy conditions, then detection capability is improved, but performance deteriorates due to noise interference
Solution Approach 1:
The audio signal is segmented into different frequency components using Mel-Frequency Cepstral Coefficients (MFCCs), which divide the frequency spectrum into mel-scale bands. This segmentation allows the system to analyze specific frequency ranges independently, enabling better discrimination of event sounds from background noise by focusing on relevant frequency segments.
Solution Approach 2:
The system transforms the audio signal from time domain to frequency domain using spectral analysis (MFCCs and Spectral Flatness Measure). This dimensional transformation allows the system to detect events based on frequency characteristics rather than just temporal patterns, providing an additional dimension for noise robustness and improving detection in noisy conditions.
2Measurement precision
If combination of different features is used to detect abnormal events, then classification accuracy is improved, but performance deteriorates due to feature redundancy and interference
Solution Approach 1:
The system merges spectral domain features (MFCCs, Spectral Flatness Measure) with time domain features (Zero Cross Rating, Short-Time Energy) into a unified feature vector. This combination leverages complementary information from both domains: spectral features capture frequency characteristics while time domain features capture temporal patterns, achieving improved classification accuracy through synergistic integration rather than redundant feature stacking.
Solution Approach 2:
The system applies parameter transformation by converting raw audio features into standardized representations (MFCCs in spectral domain, ZCR and STE in time domain). These parameter changes normalize the feature scales and distributions, enabling effective fusion of features with different characteristics while avoiding performance degradation from heterogeneous parameter spaces.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
System(s) and method(s) for classifying noise data of human crowd are disclosed. Noise data is captured from one or more sources and features are extracted by using computation techniques. The features comprise spectral domain features and time domain features. Classification models are developed by using each of the spectral domain features and the time domain features. Discriminative information with respect to the noise data is extracted by using the classification models. A performance matrix is computed for each of the classification model. The performance matrix comprises classified noise elements with respect to the noise data. Each classified noise element is associated with a classification performance score with respect to a spectral domain feature, a time domain feature, and fusion of features and scores. The classified noise elements provide the classification of the noise data.