Research On Video Background Modeling Method Based On Texture And Color Features

Video-based dynamic object detection is one of the hottest research topics in computer vision.It is a crucial issue for dynamic object tracking,recognition,and analysis.And it also has significant implications for the development of intelligent surveillance systems.While the related algorithms have made significant progress in accommodating variations in natural lighting and dynamic changes over the years,complex factors and uncertain information still significantly impact the accuracy of video object detection.To address this problem,this paper uses a mathematical,statistical background model framework,considering the connection between temporal and spatial domain information,improves the model strategy and texture features under complex factors,and combines the Dempster-Shafer Evidence Theory(DST)to construct a fast background modeling algorithm to achieve video-based motion object detection.The contents of this paper mainly include the following:(1)Research on texture and strategies in background modeling under complex scenes.To mitigate the issue of image feature sensitivity to noise and poor robustness under complex factors,we have developed a feature and improvement strategy that is suitable for background modeling algorithms.Specifically,we have designed a low-sensitivity image texture feature called Differential Local Binary Similarity Pattern with spatiotemporal information to address texture-related concerns.In terms of strategy,we have implemented a heuristic mask correction strategy and a hybrid reconstruction strategy that leverages efficient reconstruction models.Ablation experiments have demonstrated the effectiveness of these strategies in assisting algorithms to optimize false positive rates and model iteration time.(2)Texture-color background modeling algorithm involved DST.Algorithms that rely on historical information cannot fundamentally solve the uncertain information that exists in complex videos,making it difficult to ensure reliable object detection.To address this problem,DST was innovatively introduced to design background modeling algorithms,and the research mainly includes three parts.Firstly,the improved strategy and texture from the previous section are integrated to establish the algorithm framework.Secondly,the DST is applied,and the classification results are generated using the basic probability assignment method of interval numbers and the DS combination rule.Finally,the voting method is used to comprehensively consider the classification results of the DST and various features to achieve reliable detection.Experimental results on the CDnet2014 public dataset(43 complex video sequences),involving complex video data,demonstrate superior performance of the proposed algorithm under challenging scenarios,with an average F-measure of 0.789,surpassing other mathematical and statistical models.This substantiates the efficacy of the algorithm in effectively handling intricate factors and uncertain information.(3)CUDA-based parallel acceleration design.The multi-feature group bipolar model requires high computational time costs.To address this issue,parallel computing is used to improve the algorithm’s computational mechanism.Furthermore,optimizations were implemented in memory access design and I/O stream interaction using techniques such as merged memory access,page locking,and streaming thinking.Experimental results demonstrate that parallelization can improve computational efficiency by a factor of 7,thereby effectively expanding the applicability of the algorithm.