本文選題：智能視頻監(jiān)控 + 智能存儲�。� 參考：《南京理工大學(xué)》2016年碩士論文

【摘要】：隨著信息技術(shù)的快速發(fā)展,智能視頻監(jiān)控和檢索技術(shù)得到了廣泛的關(guān)注和研究。與傳統(tǒng)視頻監(jiān)控相比,智能視頻監(jiān)控和檢索系統(tǒng)涉及機(jī)器視覺、模式識別、人工智能等多個學(xué)科研究領(lǐng)域,利用智能算法使計算機(jī)協(xié)助人完成監(jiān)控和檢索的工作。本文主要設(shè)計完成的工作為：1、智能視頻監(jiān)控存儲系統(tǒng)；2、監(jiān)控視頻智能檢索系統(tǒng),該部分又分成目標(biāo)檢測、目標(biāo)跟蹤、目標(biāo)檢索三個子系統(tǒng)。文章最后設(shè)計實(shí)現(xiàn)了智能視頻監(jiān)控和檢索系統(tǒng)。傳統(tǒng)視頻監(jiān)控在連續(xù)錄像存儲過程中,存在冗余信息,既消耗大量存儲空間,又使后期信息檢索效率下降。本文研究完成了一種智能視頻監(jiān)控存儲方法,依據(jù)是否存在運(yùn)動物體,來對視頻信息進(jìn)行選擇性存儲,在節(jié)省存儲空間的同時也為后續(xù)查找和檢索提供便利。設(shè)計完成的監(jiān)控視頻智能檢索系統(tǒng),通過提取監(jiān)控視頻中的運(yùn)動物體來建立檢索庫,并根據(jù)用戶需求內(nèi)容進(jìn)行檢索。主要工作為：1、改進(jìn)的Vibe目標(biāo)檢測算法。首先通過根據(jù)前景、鬼影和各自鄰近區(qū)域直方圖相似度比較抑制鬼影；其次利用陰影亮度比背景區(qū)域低的特性去除陰影；最后根據(jù)形態(tài)學(xué)算法填補(bǔ)空洞。實(shí)驗表明相較于混合高斯建模、傳統(tǒng)Vibe算法,本文改進(jìn)Vibe算法檢測F1-measure分別提高了24%和5%,且實(shí)時性較好。2、分塊多特征自適應(yīng)融合的多目標(biāo)跟蹤算法。采用目標(biāo)底層顏色、紋理和邊緣多特征自適應(yīng)融合,對目標(biāo)和模板目標(biāo)分塊特征匹配,結(jié)合Kalman預(yù)測能夠魯棒跟蹤遮擋目標(biāo),實(shí)驗表明對遮擋目標(biāo)識別率為95.3%,每幀圖像平均處理時間為36.2ms。3、基于內(nèi)容和語義的檢索方法。結(jié)合顏色直方圖和SIFT特征進(jìn)行樣例檢索,檢索結(jié)果平均查全率為86%,平均查準(zhǔn)率為88%,平均檢索時間為13s；通過SVM對圖像分類,優(yōu)化參數(shù),語義檢索識別率為87%。最后,本文設(shè)計實(shí)現(xiàn)了智能視頻監(jiān)控和檢索系統(tǒng)。本系統(tǒng)在智能視頻監(jiān)控存儲上節(jié)省率為30%,在目標(biāo)檢測和目標(biāo)跟蹤滿足實(shí)時性的要求,在目標(biāo)檢索上查準(zhǔn)率可以達(dá)到86%以上,該系統(tǒng)能夠達(dá)到較滿意的效果。
[Abstract]:With the rapid development of information technology, intelligent video surveillance and retrieval technology has received extensive attention and research. Compared with traditional video surveillance, intelligent video surveillance and retrieval system involves many subjects such as machine vision, pattern recognition, artificial intelligence and so on. The main work of this paper is designed and completed as follows: 1, the intelligent video surveillance storage system, the intelligent video retrieval system, which is divided into three subsystems: target detection, target tracking and target retrieval. Finally, an intelligent video surveillance and retrieval system is designed and implemented. In the process of continuous video storage, the traditional video surveillance has redundant information, which not only consumes a lot of storage space, but also reduces the efficiency of information retrieval in the later stage. In this paper, an intelligent video surveillance storage method is developed, which can selectively store video information according to the existence of moving objects. It can save storage space and provide convenience for subsequent search and retrieval. The intelligent retrieval system of surveillance video is designed and completed. By extracting the moving objects from the surveillance video, the retrieval database is established, and the retrieval is carried out according to the content of the users. The main work is 1: 1, an improved Vibe target detection algorithm. Firstly, the shadow is suppressed by comparing the similarity between the ghost and their adjacent region histogram according to the foreground; secondly, the shadow is removed by using the feature that the shadow brightness is lower than that of the background region; finally, the void is filled according to the morphological algorithm. The experimental results show that compared with hybrid Gao Si modeling and traditional Vibe algorithm, the improved Vibe algorithm improves the detection F1-measure by 24% and 5% respectively, and has a better real-time performance of .2. a multi-target tracking algorithm based on block multi-feature adaptive fusion is proposed in this paper. Adaptive fusion of bottom color, texture and edge features is used to match the block feature of target and template object. Combined with Kalman prediction, the occlusion target can be tracked robustly. Experiments show that the recognition rate of occluded objects is 95.3 and the average processing time per frame is 36.2ms.3. the retrieval method based on content and semantics is presented. Combining color histogram with SIFT features, the retrieval results show that the average recall rate is 86, the average precision is 88, the average retrieval time is 13 s, and the recognition rate of semantic retrieval is 87 by SVM to classify the image and optimize the parameters. Finally, this paper designs and implements an intelligent video surveillance and retrieval system. The system can save 30% in intelligent video surveillance storage, meet the real-time requirements in target detection and target tracking, and achieve more than 86% precision in target retrieval. The system can achieve satisfactory results.
【學(xué)位授予單位】：南京理工大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2016
【分類號】：TP391.41;TN948.6
，

本文編號：1905806