Glass-break detection system using deep learning algorithm /
Burglary is one of the most common crimes throughout the world. Safety of residential places and offices is important. Researchers have pointed out that alarm systems reduce burglary rates. Existing systems mainly use magnetic reed switches to detect unauthorized openings of doors and windows. Glass...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Subjects: | |
| Online Access: | Click here to view 1st 24 pages of the thesis. Members can view fulltext at the specified PCs in the library. |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Burglary is one of the most common crimes throughout the world. Safety of residential places and offices is important. Researchers have pointed out that alarm systems reduce burglary rates. Existing systems mainly use magnetic reed switches to detect unauthorized openings of doors and windows. Glass panels are very common in homes and commercial buildings. Despite the use of iron grilles in conjunction with glass panels burglary happens, iron grilles can easily be removed by unscrewing. A burglar would then just shatter the glass and come in freely without ever tripping an alarm. This spawns interest in glass-break detectors. State-of-the-art glass-break detectors simply threshold sound levels to determine whether the acquired signal is loud enough to be a glass-break signal. This causes many false alarms since there are many types of loud sounds. Even with state-of-the-art digital signal processing techniques, researchers have not been able to successfully construct the accurate model of glass-break signal since there are infinitesimal variations of them. Moreover, certain sounds such as thunder sounds, shouting, gunshot, hitting objects are similar in frequency domain to that of sounds arising from glass breaking events. The author proposes a three-layered framework to perform real-time detection of glass breaking sounds using an end-to-end deep learning approach. The first layer performs background subtraction of audio from a noisy environment using a predefined average energy threshold technique and converts raw audio into an image representation. The second layer performs end-to-end feature learning and detection of glass breaking sounds based on image representation of data using deep convolutional neural network with transfer learning approach. The final layer performs online learning (automatic retraining) by employing decentralized federated online learning approach. Experiments have been conducted both offline and in real time. The results demonstrate the proposed glass-break detection approach achieved over 99.3% detection accuracy with significantly lower false alarm rate than that of the state-of-the-art systems. The limitation of the proposed system is that at the moment it only works well with normal single-pane non-tempered glass types with less than 20mm thickness. However, the proposed system has the learning-based subsystem with the potential to be trained to work with all glass types by simply providing enough samples. |
|---|---|
| Physical Description: | xix, 176 leaves : illustrations ; 30cm. |
| Bibliography: | Includes bibliographical references (leaves 142-153). |