A forest pyrotechnics detection method based on 3D convolution neural network

Abstract

The invention provides a forest pyrotechnic detection method based on 3D convolution neural network, which comprises the following steps: establishing a preliminary forest pyrotechnic image sample data set; the structure of neural network based on 3D convolution is designed, and the time-depth parameters of convolution and pooling operation are introduced; performing training and optimization on 3D Convolution neural Network with the data set; the video to be processed is motion detected in blocks, and the suspected smoke area is judged. Using the optimized 3D convolution neural network modelto perform forward computation, the classification result of the suspected smoke region is obtained. The invention is based on 3D convolutional neural network, through 3D convolutional kernel slidingon time axis to combine spatial features of continuous frames, mining spatio-temporal features of continuous sub-images, ensuring high detection rate and reducing false positives caused by a large number of single-frame images, thereby greatly improving identification accuracy and possessing better robustness and generalization ability.

Description

technical field [0001] The invention belongs to the field of computer vision, and relates to a video monitoring method of forest fireworks, in particular to a method for detecting forest fireworks based on a 3D convolutional neural network. Background technique [0002] Forest fire refers to the behavior of forest fire that loses human control, spreads in a large range of forest areas, and causes a lot of losses to forest ecosystems. With the warming of the climate and frequent occurrence of extreme weather, the world has entered a period of high incidence of forest fires, and the risk of forest fires has increased. In 2016, there were 2034 forest fires in my country, an increase of 30.72% over 2015. Among them, there was one major fire, which increased by 83.33% in 2016 compared to 2015. In 2016, the total area of ​​the fire site was 18,161 hectares, of which 6,224 hectares were affected forests. In 2016, my country's fire fighting funds were 120.1739 million yuan, an in...

AI Technical Summary

Problems solved by technology

However, most of the existing deep learning-based methods cannot learn the motion information of sequence images. If the general-purpose convolutional neural network method is used directly, the effect is usually not ideal, because a lot of interference information is usually introduced in forest fire prevention video surveillance. , such as birds, swaying branches, mosquitoes, monitoring pan-tilt shaking, etc., will interfere with the accuracy of the algorithm to a certain extent; the scene in the monitoring is usually more complicated, and shadows, vehicles on the road, and shaking treetops are easy to cause False detection

Traditional manual features cannot adapt to various scenarios

Although the convolutional neural network has powerful feature extraction capabilities, it is still prone to misjudgment of smoke for some mist and roads. This is because the 2D convolutional neural network only extracts the static texture features of the image, but the human eye recognizes the smoke When most scenes are dependent on the motion characteristics of the smoke, especially when the target is far away

If it is only based on a single image, many accidental factors will be introduced, resulting in an increase in the false positive rate

Therefore, the existing detection faces the problem of high false alarm rate in practical applications, and cannot cope with a variety of complex scenarios.

Images