Research On Remote Monitoring Technology For Wildlife Conservation

Wildlife is an important component of the natural ecosystem,and protecting wildlife is extremely important for maintaining ecological balance and biodiversity.Through infrared cameras and acoustic devices,images and acoustic data of wildlife can be collected remotely,and deep learning technology can be used to automatically process and analyze data,making wildlife protection more targeted.Networked infrared cameras enable real-time transmission of wildlife monitoring data over the network,however,the deployment of cameras did not consider the signal coverage of the monitoring area,and there are challenges in automated analysis of monitoring data,including low precision of animal recognition in nighttime infrared scenes,and the need for animal localization combined with species classification for biodiversity monitoring.Therefore,this thesis focuses on three aspects: networked infrared camera deployment,animal recognition in nighttime infrared scenes,and animal sound event detection and localization.Main research contents are as follows:1.Particle swarm-based method for remote monitoring of wildlife infrared cameras.The signal propagation model and particle swarm algorithm are combined to solve the networked infrared camera deployment location,so that it meets the monitoring requirements within the network coverage.In the iterative process of deployment location solution,the algorithm is optimized from three aspects: learning weights,the worst fitness particle location and the optimal fitness particle location.The experiments show that the search capability of the algorithm is improved,and the coverage increases by 16% with the same number of cameras after the algorithm converges.2.Animal recognition in nighttime infrared scenes based on generative adversarial network domain migration.For the poor visual saliency and problem of low recognition rate of nighttime images of animals,cyclic generative adversarial network is used to migrate infrared images of animals into visible images.Through multi-scale convolution,this thesis constructs a model named MCCGAN for infrared image domain migration,extract features of different fineness,and make the generated visible images have better overall quality and visual effect.The resulting visible images have better overall quality as well as visual effects.The experiments show that MCCGAN can effectively migrate the infrared domain images of animals to the visible domain images,and the average recognition rate of animals in infrared scenes is improved by 7.5%.3.Animal sound event detection and localization based on residual attention network.With convolutional recurrent neural network struct,target animal sound fragments are detected and species identification and location estimation are performed.By residual network and attention mechanism,the model’s ability to extract sound features is improved,and deep separable convolution is used to improve its computational speed and reduce the number of parameters of the model.Experiments show that the proposed model can achieve the detection and location estimation of target animal sound fragments,and the detection accuracy is improved by 11.4% and the localization error is reduced by 0.81 m compared results without residual attention network.