Dual Deep Learning Model for Image Based Smoke Detection

• Published in: Fire technology 2019-11, Vol.55 (6), p.2419-2442
• Main Authors: Pundir, Arun Singh, Raman, Balasubramanian
• Format: Article
• Language: English
• Subjects: Alarm systems; Algorithms; Artificial neural networks; Cameras; Characterization and Evaluation of Materials; Civil Engineering; Classical Mechanics; Deep learning; Detectors; Edge detection; Engineering; Feature extraction; Fog; Frames; Image detection; Motion detection; Neural networks; Object recognition; Optical flow (image analysis); Pattern recognition; Physics; Robustness; Sensors; Smoke; Support vector machines; Wildfires
• ISSN: 0015-2684; 1572-8099
• DOI: 10.1007/s10694-019-00872-2

Image-based smoke detection could help in faster and robust detection and monitoring of wildfires. It is becoming the best alternate of sensor based detectors for early detection of wildfire. The limitations of sensor based detector is that, they need close vicinity to fire for raising the alarm which make them vulnerable in case of detecting far-distant wild fire. Hence, vision based detection system which utilizes the surveillance cameras which shows more fastness and robustness as compared to sensor based detectors. These cameras when installed on hill top or mobile tower can raise the early alarm for any possibility of smoke present in the frames of videos whether near-by or far-away smoke. The proposed work presents a robust method for smoke detection which, utilizes a dual deep learning framework. The proposed architecture makes use of framework based on Deep Convolutional Neural Networks, which has proven their supremacy in object recognition tasks. The first deep learning framework is employed for extracting the image-based features from smoke patches, which are being extracted using superpixel algorithm. We have employed total of 20,000 frames with equally distribution of non smoke and smoke classes, out of which 6000 frames are utilized for testing purpose and 14,000 are used for fine tuning purpose. These features are comprised of smoke-color, smoke-texture, sharp edge detection and perimeter disorder analysis. The second deep learning framework is used for extracting motion-based features such as moving region of smoke, growing region and rising region detection. Optical flow method is employed, in order to capture the random motion of smoke. These extracted optical flow are then feed into Deep CNN for extracting motion based features. Features from both the framework are combined to train the Support Vector Machine and end to end classification which is CNN classifier. Accuracy on the nearby smoke and faraway smoke is 98.29 % and 91.96 % respectively. Testing on different varieties of non-smoke videos such as clouds, fog, sandstorm and images of cloud on water, method proves its precision and robustness. The average accuracy in all the scenarios is 97.49 % which outperforms the state of the art method for these scenarios. Contribution of this work lies in the fact that we have given 20,000 frames based smoke and non-smoke dataset and secondly our method outperform the existing method on challenging imaging conditions.