Farmland Video Moving Target Detection Based On Spatiotemporal Saliency Fusion

With the development of China’s agriculture,the issue of farmland economic crop safety is becoming more and more important,but the farmland environment is complex,the dynamic background is discontinuous,and there are high-frequency disturbances such as swaying of the vegetation,which can easily lead to missed detection and false detection.In addition to the environmental factors that are susceptible to illumination and irregular swaying,the farmland scenes increase the detection difficulty in the large-scale screen of the surveillance video.Many moving target detection algorithms work well in simple scenarios,but In the farmland scene,the false detection rate is high,and the moving target detection is an important basis for the tracking of moving targets.Therefore,it is necessary to conduct in-depth research on the moving target detection of farmland video.In this paper,the time domain based moving target detection method and visual attention based spatial saliency detection are studied.Based on this,a method of detecting temporal and spatial variability of farmland video moving target is studied and proposed.The main research contents and conclusions of this paper are presented.as follows:(1)Research on moving target detection method based on time domainIn view of the current high false detection rate of farmland video moving target detection,firstly,three kinds of methods based on time domain for moving target detection are studied,and two specific background difference algorithms are studied,including researching mixed Gaussian modeling method for Subsequent comparison experiments,the visual background extraction algorithm is studied and its advantages and disadvantages are analyzed for subsequent improvement.The visual background extraction algorithm has superior performance,but it is easy to generate noise and holes,which can be improved to obtain time saliency map.(2)Spatial saliency detection based on visual attentionAiming at the shortcomings of time-domain based moving target detection methods,the visual attention mechanism and spatial saliency detection methods are studied from the perspective of space,and the spatial saliency detection methods of different principles are studied,mainly based on scene perception,histogram comparison and dense sparse reconstruction.Methods.Research and propose a method of intensive sparse reconstruction for farmland spatial significance detection,The experimental results show that the dense sparse reconstruction method can obtain a better spatial saliency map in the farmland environment.(3)Farmland video moving target detection method based on spatiotemporal saliency fusionBased on the research of time domain moving target detection and spatial saliency detection,this paper studies and proposes a method of farmland video moving target detection with spatiotemporal saliency fusion.Firstly,based on the spatial sparsity of the dense sparse reconstruction method,combined with spatial saliency to improve the visual background extraction algorithm to calculate the time saliency,then Bayesian fusion method for space-time significant fusion and finally extract the moving target..The experimental results show that the proposed moving target detection algorithm exhibits high robustness in complex scenes with challenging interference factors.In the five groups of video used for detection,it involves swaying of the leaves,slight shaking of the camera,illumination,etc.The experimental results were analyzed from the perspectives of effect comparison,accuracy of moving target detection and integrity of foreground target detection.The average accuracy of the final detection reached 92.15%,and the average recall rate reached90.75%,which was superior to other similar algorithms.