Aim:
Ā Ā Ā Ā Ā Ā Ā The aim of this study is to address the escalating issue of wildfires on a global scale, particularly in regions like Brazil, where the Amazon forest and other forest biomes are significantly affected. The proposed aim is to introduce a novel and lightweight convolutional neural network (CNN) model for the real-time detection of wildfires using RGB images. This approach seeks to overcome the limitations of existing methods employed for wildfire detection and offer enhanced advantages. The primary objective is to create a CNN architecture that can effectively process images from various sources, including unmanned aerial vehicles and video surveillance systems, by integrating edge computing devices for image processing.
Abstract:
Ā Ā Ā Ā Ā Ā Ā Forest fires may cause serious human and economic losses, and it is also a key technology linked to realizing forest intelligent fire detection and prevention. Therefore, the study of forest fire detection and early warning system has great social application value. However, the traditional methods of detecting fires have weak real-time and detection capabilities. The new generation of deep learning technology, especially convolutional neural networks, provides new means and methods for fire detection. The convolutional neural network has the characteristics of good fault tolerance, self-adaptability, self-learning ability, and weight sharing, which makes the convolutional neural network can identify fires in a working principle close to that of the human eye. The fire detection method based on deep learning has high accuracy and robustness. However, due to its large number of parameters and a large number of calculations, it is difficult to be practically applied. With the development of edge computing, it is now possible to use the edge instead of the cloud for computing, which will greatly reduce latency. This paper collects a large number of forest fire images as a data set, trains a model that can recognize forest fires based on the YOLOv7 network, and further deploys the trained model to the edge server RK3588. Finally, a front-end visual interface capable of displaying fire recognition results was built using Pyqt5. The model in this method has good robustness and generalization ability, and the application of edge computing in fire forest detection further improves the speed of fire detection, which provides a new way for forest fire detection.
Introduction:
Ā Ā Ā Ā Ā Ā Ā Fires have become more frequent in recent years. Fire prevention and detection is an important research project which is beneficial to national economy, peopleās life safety and natural environment. Traditional flame monitoring mainly uses smoke sensors and temperature sensors. Traditional flame monitoring is limited to a fixed and closed small space and relies on monitoring the smoke concentration and temperature threshold in the closed space to detect. The ability to detect single flame is limited. At the same time, due to the limitation of space, the conditions of outdoor and spatial-temporal flame detection cannot be met. In view of this, Lasaponara et al. proposed an improved adaptive flame detection algorithm based on AVHRR (Advanced Very High Resolution Radiometer); Celik et al. proposed a real-time flame detection algorithm that combines target foreground information with color pixel statistics; Zhou et al. proposed a flame detection based on flame contour determine whether a target is a flame target based on three features: contour area, edge, and roundness of the detected target. Because the flame target features are affected by color, contour changes and complex scenes, the traditional flame target detection is prone to false detection and the problem of missing detection for small-sized targets. Compared with the traditional flame detection methods, the exposed detection conditions are limited, the detection method is single, and the detection performance is worse. In the past decade, deep learning-based flame detection methods have developed rapidly. In the literature, a small-scale flame detection method based on YOLOv7 algorithm was proposed to achieve the detection of different scales of flames using an improved clustering algorithm. In the literature a Fire-YOLO algorithm is introduced, which adds depth-separable convolution to reduces the computational and parametric quantities of the model, and improves the perceptual field of the feature layer by using cavity convolution, and achieves a detection speed of 4 frames/sec. A new adaptive selection algorithm for flame image features is introduced in the literature, which introduces genetic optimization to the attribute approximation of rough sets and increases the diversity of the population by dynamically pruning and supplementing new individuals, effectively improving the generalization ability of the flame recognition algorithm. The current advanced target detection methods are mainly single-stage and two-stage algorithms. For example, single-stage detection algorithms: YOLO The single-stage detection algorithm, compared to the two-stage detection algorithm, has the property of fast detection speed, which can better meet the real-time flame image detection. Based on this, a single-stage detection algorithm is preferred. The YOLOV7 detection method is proposed on the basis of the most advanced target detection method YOLOv7. By combining ConvNext Block to build the CN-B network module, replace the first and last ELAN module of Backbone in YOLOv7, Replace the Bags module (Trainable bag-of-freebies in YOLOv7), the ELAN variant of Head, with the CN-B network module. The YOLOv7 network model is not only lightweight, but also enables the YOLOv7 network to obtain larger sensitivity field, enhance the ability of flame feature extraction, improve the network performance, obtain higher accuracy and mAP, achieve smaller Parameters, Gradients, Layers and less computation.
Proposed System
Ā Ā Ā Ā Ā Ā Ā Ā YOLOv7 uses Model scaling for concatenation-based models. When a cascade-based model performs depth scaling, the output width of the computational block also increases. This phenomenon will lead to an increase in the input width of the subsequent transport layers. Therefore, it is proposed that when performing model scaling for cascade-based models, only the depth in the computational block needs to be scaled and the remaining part of the transport layer is performed using the corresponding width scaling. Also, the parametric convolution is re-analyzed by using the gradient flow propagation path in combination with different networks. after analyzing the combination and corresponding performance of RepConv with different architectures, used the RepConv without identity connectionĀ to design the architecture of the planned re-parametric convolution, avoiding that the RepConv that the identity connection breaks the cascade of residuals, thus providing more gradient diversity for different feature mappings. YOLOv7 proposes a new label assignment method which guides the auxiliary and bootstrap heads by bootstrap head prediction. In other words, the bootstrap head prediction is used as a guide to generate hierarchical labels from coarse to fine, which are used for auxiliary head and bootstrap head learning, respectively.
Advantages:
Ā Ā Ā Ā Ā Ā Ā Compared to earlier versions, YOLOv7 offers improved accuracy, speed, and versatility. Its architecture enables efficient detection of objects like fires in images, making it an ideal candidate for real-time forest fire detection.
Reviews
There are no reviews yet.