Forest Fire Detection In Chongli Region Based On Video Images Of Unmanned Aerial Vehicle

Forests are strategic resources to sustain China’s sustainable economic development and play an essential role in social,cultural,and ecological aspects.Forest fire is highly destructive and covers a wide area.It affects the local environment,threatening residents’ life and property safety.With the impact of global climate change and human production activities,fire disasters have occurred frequently in recent years,and then all countries actively conduct fire prevention and control research.However,monitoring early fires in forest areas is challenging because of fires’ strong burst and concealment.Therefore,real-time detection of suspicious fire and smoke in forest areas is beneficial to finding and extinguishing fire sources in time and plays a positive role in protecting forest resources.This paper takes forest video data from the perspective of UAV(Unmanned Aerial Vehicle)in Chongli District as the research object and carries out research on forest fire target detection based on vision.The main contents are as follows:1.Data processing and cleaning.To carry out forest fire research from the perspective of UAVs,people need to solve data collection and related labeling problems.The specific work is as follows: 1)In different seasons,we went to Chongli District of Hebei Province for field sampling of raw data multiple times and used UAVs to collect artificially simulated data in various scenarios.The data were recorded in the form of video,and then forest fire detection objects were determined according to forest flames and smoke characteristics;2)Formulate detailed data processing procedures and provide corresponding reference annotation cases for different types of samples;3)In order to further improve the processing efficiency of video data,semi-automatic annotation method based on algorithm is adopted;4)The final sample number is 6032,and the data includes different scenes in multiple seasons.2.A data enhancement method based on adaptive sample equalization and information fusion was proposed to alleviate problems such as insufficient forest fire samples,unbalanced type distribution,and insufficient scene expression ability.I analyzed the Chongli data set was analyzed by category statistics,annotation centralization,and target normalization.The random data enhancement,Mosaic,and Cutmix enhancement methods were theoretically explained,their advantages and disadvantages were extracted,and a self-adaption Mix Augmentation(SMA)enhancement method was proposed combined with the results of data analysis.Sixteen comparative experiments were designed,and mAP was used as the evaluation index to compare the effects of traditional data enhancement,Mosaic,and the method proposed in different algorithms.The results showed that the proposed method improved 11.95% and 4.76% in SSD compared with the random data enhancement and Mosaic methods,respectively.In the YOLOv3,the mAP increased by 16.33% and 1.06%.In the YOLOv4 model,they increased by 18.24% and1.79%,respectively.In YOLOv5,it increased by 3.10% and 0.65%,separately.3.A lightweight Parallel-YOLO target detection algorithm is proposed.According to YOLOv5 official source code,analyzing the overall structure of the model,feature extraction module,and loss function design.Integrate Chongli and FLAME open data to construct a small target forest fire data set for the experiment.Based on HRNet and YOLOv5s,this paper designs a Parallel YOLO algorithm that is lighter than YOLOv5s.The structure,feature extraction module,multi-level resolution information exchange layer,and fusion layer of the network are introduced in detail.The version based on the channel attention mechanism is given.In order to verify the effectiveness of the algorithm,a comparative experiment was designed to analyze the AP50,parameter number,and FPS of YOLOv5s and Parallel-YOLO at different resolutions.The results show that in the input resolution of 128 samples,the Parallel-YOLO proposed in this paper improves by 1% compared with YOLOv5.The algorithm using the Fusion-SE module is 4% better than YOLOv5s.Taking FPS as an indicator,the maximum frame of Parallel-YOLO was 117 frames,and we can see there are 9 frames higher than YOLOv5s.Compared with YOLOv5s,parameters decreased by more than 16%.4.Designed and developed a system for forest fire detection and warning based on UAVs combined with the algorithm.Designed fire protection system schemes according to project requirements and basic site information.The problem of response delay of forest fire prevention systems based on UAVs is discussed.At the same time,the specific assumptions and data collected in the field are given.The final calculated result is 6 seconds delay.I also introduced the functional requirements of each module,the corresponding implementation technology,and the cooperation logic between modules.Finally,the system’s interface is introduced in detail,and the simulation test effect is given.