Spectrogram Directional Filtering for Falling Action Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current action detection technologies face challenges in accurately distinguishing falling actions from non-falling actions due to similarities in spectrogram energy distribution, leading to false alarms, and are hindered by unevenly distributed training data, which reduces the model's ability to learn falling features effectively.
Innovation Solution
A feature enhancement and data augmentation method that includes directional filtering and data augmentation algorithms to enhance falling action features in spectrograms, improving discernment and increasing the amount of falling data, thereby strengthening model generalization capability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional spectrogram analysis is used for falling action detection, then the detection system can operate with simple processing, but the model generates false alarms due to similar energy distribution between falling and non-falling actions
Solution Approach 1:
The spectrogram is segmented into multiple directional components using directional filtering. The filtering process divides the spectrogram into different directional segments (horizontal, vertical, diagonal) to separately analyze energy distribution patterns in each direction, enabling the model to distinguish falling actions from non-falling actions based on directional energy characteristics
Solution Approach 2:
The patent introduces a directional dimension to the traditional spectrogram analysis by applying directional filtering at multiple angles (0°, 45°, 90°, 135°). This transforms the two-dimensional spectrogram into multi-dimensional directional energy distributions, providing additional discriminative features for action classification
2Reliability
If actual field data is collected for training, then the model learns from real scenarios, but the training data is unevenly distributed with far fewer falling actions than non-falling actions
Solution Approach 1:
The patent creates synthetic falling action samples by copying and transforming existing falling action data through data augmentation techniques. These synthetic samples are generated by applying various transformations (noise addition, time shifting, frequency modulation) to original falling action spectrograms, thereby increasing the quantity of training data without requiring additional real-world falling events
Solution Approach 2:
The patent applies parameter changes to existing training data through data augmentation. By modifying parameters such as noise levels, time offsets, frequency shifts, and energy distributions, the system generates diverse variations of falling and non-falling action spectrograms, creating a more balanced and robust training dataset
Data Source
AI summary
A feature enhancement and data augmentation method for detecting at least one action of at least one tested subject is provided. The feature enhancement and data augmentation method includes obtaining at least one spectrogram; and performing a feature enhancement processing on the at least one spectrogram, to enhance at least one feature corresponding to at least one action in the at least one spectrogram and generate at least one feature enhanced spectrogram. The feature enhancement processing comprises a directional filtering.