Research On Video-based Wild Animal Object Detection Algorithm

Biological resources are the natural basis for human survival and development,and a powerful guarantee for the balance and stability of the ecosystem.Wildlife resources are an important part of biological resources.Conservation and rational utilization of wildlife resources are of great significance for sustainable development.However,the current monitoring and protection of wildlife is through on-the-spot exploration,or the use of expensive real-time video recorders for uninterrupted fixed-point video recording,which requires a lot of manpower and material resources.Video-based wildlife target detection method can automatically detect the type and location of wildlife,and can enhance the protection of wildlife.In the process of target detection for wildlife video,the stability and accuracy of target detection will be reduced due to the influence of motion blurring,deformation,illumination and occlusion.The solution of occlusion problem in target detection is the key to improve the stability and accuracy of target detection,so we will study the occlusion problem in video target detection.Aiming at the problem of wildlife occlusion in video,this paper first studies a series of algorithms of target detection based on deep learning,and then deeply studies YOLOv3(You Only Look Once)target detection algorithm.Taking wildlife video as the research object,starting from the unique time series relationship of video,two algorithms are modified on the basis of YOLOv3 algorithm.Advance and experiment.The main contents of this paper are as follows:(1)A video-based wildlife Video Detection Datasets(WVDDS)is constructed.The WVDDS video data is labeled manually according to the frequency of every five frames.The format of the WVDDS is PASCCAL VOC format,which contains 12 categories and a total of 6601 video images.(2)A detection method of wildlife video object occlusion based on entropy and motion offset is studied.This method uses Darknet-53 model to extract target features;adds additional convolution to the Darknet-53 model to detect the target for the first time;and for the failed image,obtains the detection position and category of the new target according to the image detection information corresponding to the position offset and the lowest information entropy in the video sequence;finally,it is tested on the WVDDS data set.The results were compared and analyzed.The experimental results show that the modelcan detect occluded objects effectively,but the detection results of the current frame depend on the detection results of the preceding frame,and when the target moves from the moving state to the stationary state and the stationary state remains for a long time,a certain degree of position offset will occur.(3)Aiming at the shortcomings of wildlife video object occlusion detection method based on entropy and motion offset,a wildlife video object detection method based on multi-feature image fusion is proposed in this paper.The method uses mutual information entropy to fit the correlation factors of frame fusion,uses correlation factors to iterate the feature maps of three different scales extracted from the front and back frames of video,and uses histogram equalization to calculate similarity to judge the critical condition of feature map fusion.The experimental results show that the method does not depend directly on the detection results of the preamble frame,and does not produce position offset.Moreover,it can solve the problem of video occlusion well,and the accuracy of target detection has been improved.