Advancing Shooter Detection for Public Safety

Gun violence is a major issue in the United States. To address this problem, there is a growing effort to develop systems that can help improve public safety, specifically by detecting and tracking shooters. The goal is to understand where shooters are and what they are doing, which can help prevent or lessen the harm caused by violent events.

#Current Approaches

Most existing systems focus on detecting guns specifically. However, our approach is different: we want to detect the shooter as a whole person, not just their weapon. This way, even if the gun is hidden from view, the system can still identify the shooter. Unfortunately, there isn't much public data available on shooters, making this task more challenging.

To tackle this issue, we turn to synthetic data, which is computer-generated information that mimics real-world scenarios. We used tools like Unreal Engine to create virtual environments where shooters and others interact. By training our system with this synthetic data, we hope to improve its ability to work well in various situations.

#Methodology

#Creating Synthetic Data

We created synthetic training data using Unreal Engine, which allows us to simulate different environments. The virtual scenarios included places like schools, hospitals, and shopping malls. In these simulations, we programmed actors to act like shooters and people trying to escape. We then captured video footage of these simulated events to use for training our detection and tracking system.

#Domain Randomization

One challenge with synthetic data is that it doesn't always translate well to real-world situations. To improve this, we used a technique called domain randomization. This means we made our synthetic data more varied to help the system learn better. For instance, we changed the colors of objects and the positions of actors in the simulations to create many different scenarios.

#Training the Model

We used a popular detection model called YOLOv8 for our system. This model is efficient and effective, especially for detecting smaller objects like guns. We trained it using various combinations of real and synthetic data, testing which mix gave us the best results. During the training, we focused on two main classes: shooters and guns.

#Detecting and Tracking Shooters

Once our model was trained, we needed a way to track shooters effectively. We used a tracking system called Deep OC-SORT combined with another component called OSNET for identifying individuals. This allowed us to follow shooters over time, even when they might be obscured or hidden from view.

#Gun Detection Confirmation

To ensure our tracking system is accurate, we implemented a method where the presence of a gun is used to confirm that a detected individual is a shooter. In other words, before we label someone as a shooter, we check if they are holding a gun.

#System Performance Evaluation

After developing our system, we needed to assess how well it actually works. We did this by evaluating both detection and tracking performance using various metrics. These evaluations were performed using real videos, some of which did not feature any shooters to test the system's ability to avoid false alerts.

#Results

#Detection Performance

We tested multiple versions of our detection model, training each one on different combinations of real and synthetic data. The results showed that using both types of data enables the model to perform better than using real data alone.

#Tracking Performance

We also evaluated the tracking performance of our system. This involved analyzing how well it maintained the identity of a shooter over multiple video frames. We noticed that there were instances where the system struggled with consistently identifying the same individual, which is known as ID switching.

#Edge Device Functionality

One of the key goals of our system is to run on low-cost computing devices, known as edge devices, such as the Raspberry Pi and Jetson Nano. We checked how quickly our system can process information on these devices, ensuring it can deliver real-time alerts in critical situations.

#Addressing Challenges

While we made significant progress, several challenges still remain. The detection of guns tends to be less accurate than detecting shooters. This is likely due to the guns being smaller and less distinguishable in different settings. There is also a risk of privacy concerns since our system relies on video footage captured by existing security cameras.

#Future Directions

Moving forward, there is a clear opportunity to enhance our system's performance by improving gun detection capabilities. This could involve integrating our findings with existing gun detection datasets to create a more robust model.

Additionally, further testing in real-life settings will help refine our approach, ensuring it can adapt to the complexities of various environments and situations. Exploring other technologies for improved tracking and detection can also be beneficial.

#Conclusion

The fight against gun violence requires innovative solutions. Our detection and tracking system aims to enhance public safety by providing quicker and more accurate information about potential threats. By harnessing the power of synthetic data and advanced tracking methods, we hope to contribute to creating safer public spaces.

As we continue to refine our technology, we are committed to addressing challenges while prioritizing the privacy and safety of individuals. Our goal is to not only improve detection and tracking methods but also facilitate effective response strategies in emergency situations.