Industrial smoke detection systems that use real-time surveillance video often have a hard time seeing smoke when the environment is hazy due to fog or snow. This can be a significant problem when it comes to disaster management. To address this, a team from Sejong University in Seoul, South Korea created an edge intelligence-assisted method to detect smoke with fewer false alarms.
Current video-based smoke detection systems rely on color, motion, or a hybrid of the two to identify smoke. These systems often result in false alarms and have limited accuracy in hazy environments or at greater distances. They also tend to require extensive computer processing power.
“Existing smoke detection methods mistake fog or other weather abnormalities for smoke due to their similar color and patterns,” said Khan Muhammad, a professor at Sejong University.
Muhammed and his team created a detection method that is both more accurate and more energy efficient. Using edge intelligence, the software can be deployed over resource-constrained devices to make decisions without having to send data to the cloud or a centralized server. The lightweight convolutional neural network architecture can be deployed over the types of resource-constrained devices (Raspberry pi or FPGA) used in current disaster management systems.
Figure 1: A depiction of the edge intelligence-assisted smoke detection system for foggy environments.
“We focused on two main problems: edge computing and smoke detection in a hazy or foggy environment,” said Muhammad. The goal was to create a system that could be easily integrated into existing surveillance camera technologies.
To test their system, the team created three datasets with four different types of video, including smoke in a clear environment, no smoke, smoke with fog, and no smoke with fog, as shown in Figure 2 below. The new method was accurate 98.17% of the time, with a false-alarm rate of just 1.18%. The team compared its system to three other smoke detection systems, which had false-alarm rates of 2.01 – 4.16%.
Figure 2: The smoke detection system was tested in foggy and clear environments to determine how frequently it reported a false alarm.
The team is now working on ways to extract contextual information from the environment, so that its system can analyze the entire scene and not just the smoke. The new fire detection system could be a good fit for smart cities that use surveillance cameras to detect various anomalous events, Muhammed said. It also could be integrated into existing home surveillance cameras.
“Smoke is often the initial sign of a fire, and its early detection can minimize significant losses for any industry,” said Muhammad.
For more information on edge computing, visit the Xplore digital library.
