Researchers recorded gunshot-like sounds in places where guns might be fired (including outdoor parks).
According to the gun violence file, there were 296 mass shootings in the United States this year. Sadly, 2021 will be the deadliest year of gun violence in the United States in the past 20 years.
Distinguishing between dangerous audio events (such as shooting a gun) and non-life-threatening events (such as a plastic bag burst) can mean the difference between life and death. In addition, it can also decide whether to deploy public safety personnel. Humans and computers often confuse plastic bag popping with real gunshots.
In the past few years, there has been a certain degree of hesitation in the implementation of some well-known available acoustic gunfire detector systems because they can be costly and generally unreliable.
In an experimental study, researchers from the School of Engineering and Computer Science at Florida Atlantic University focused on solving the reliability problems related to the false alarm rate of these detection systems. The model’s ability to correctly recognize sounds can distinguish well-trained models from inefficient models even in the most subtle scenarios.
Considering all the arduous tasks of gunshot-like sounds, the researchers created a new data set that includes recordings of plastic bag explosions collected under various environments and conditions, such as the size of the plastic bag and the distance from the recording microphone. The recording duration of an audio clip is 400 to 600 milliseconds.
The researchers also developed a classification algorithm based on Convolutional Neural Networks (CNN) as a baseline to illustrate the relevance of this data collection effort. These data are then used with the gunshot data set to train a CNN-based classification model to distinguish life-threatening shooting incidents from non-life-threatening plastic bag explosions.
The results of the study, published in the Sensors magazine, show how fake gunshots can easily confuse gunshot detection systems. 75% of plastic bag blasts are incorrectly classified as gunfire. The deep learning-based classification model trained using the popular city sound dataset containing gunshots cannot distinguish between plastic bag popping and gunshots. However, once the plastic bag popping sound was injected into the model training, the researchers found that the CNN classification model performed well in distinguishing actual gunfire from plastic bag sound.
"As humans, we use additional sensory input and past experience to recognize sounds. On the other hand, computers are trained to decipher information that is usually unrelated or imperceptible to the human ear," senior author, Department of Electrical Engineering and Computer Science, School of Engineering, Professor and Dean Dr. Zhuang Hanqi said. And computer science. "Similar to how bats dive around objects, because they emit high-pitched sound waves at different time intervals, and then we use different environments to allow machine learning algorithms to better perceive the differences in closely related sounds."
In this study, gunshot-like sounds were recorded at locations where gunshots were possible, including a total of eight indoor and outdoor locations. The data collection process started with experiments on various types of bags, and the trash can lining was selected as the most suitable. Most audio clips were captured using six recording devices. To check the extent to which the sound classification model may be confused by fake gunshots, the researchers trained the model without exposing it to the sound of plastic bag popping.
Initially, there were 374 gunshot samples used to train the model. These samples were obtained from the city sound database. The researchers used 10 categories in the database (gunshots, dog barking, children playing, car horns, air conditioning, street music, sirens, engine idling, jackhammers, and drilling). After training, the model was then used to test its ability to exclude plastic bag popping as real gunshots.
“The high percentage of misclassifications indicates that the classification model has difficulty distinguishing gunshot-like sounds, such as those from plastic bag popping and real gunshots,” said lead author and doctoral student Rajesh Baliram Singh. FAU student in the Department of Electrical Engineering and Computer Science. "This guarantees the process of developing a data set that contains sounds similar to real gunshots."
In gunshot detection, having a database of specific sounds that may be confused with gunshots but with a rich diversity can lead to a more effective gunshot detection system. This concept prompted researchers to create a database of plastic bag explosion sounds. The higher the diversity of the same voice, the greater the probability that the machine learning algorithm will correctly detect that specific voice.
“Improving the performance of the gunshot detection algorithm, especially reducing the false alarm rate, will reduce the chance of treating harmless audio trigger events as dangerous audio events involving firearms,” said Dr. Stella Batalama, Dean of Engineering and Computer Science. "This data set developed by our researchers, together with their classification models trained on gunshots and gunshot-like sounds, is an important step in reducing false alarms and improving overall public safety by deploying key personnel only when necessary."
The co-author of the study is Jeet Kiran Pawani, MS, who conducted the study while at Georgia Institute of Technology.
Researchers use an anechoic room as one of the environments, which provides "pure" and undisturbed samples, adds a lot of information to the CNN, and makes the model more robust.
Tags: Technology | Students | Faculty and Staff | Research | Engineering