Content based video retrieval using dynamic textures

B. Mounika

doi:10.1007/s11042-022-13086-6

Content based video retrieval using dynamic textures

2022-06-03

2

PRE

AI

Hotta Himanshu Sekhar

K

Khare, Ashish

DOI：10.1007/s11042-022-13086-6

原文链接

原文求助

摘要

En 中文

In recent days, design of smart city applications is attracting several researchers offering improvised services to citizens through efficient management of day to day life activities such as business, safety, public utility, transportation and hospitality etc. In most of the smart city applications video retrieval plays crucial role. Through implementation of video retrieval we can achieve several smart city innovates such as monitoring of traffic or crowd etc. The existing video retrieval system depends on either ontological concepts or training based concepts. Implementation of ontological concepts has shown drawbacks such as miss-assignment of tags to the database videos leads to poor efficiency and also needs huge manual effort for the purpose of onnotation. Training based concepts needs prior knowledge and also takes more time for the purpose of training and to overcome all these drawbacks content based video retrieval systems (CBVR) has been evolved. In most of the existing CBVR systems the principal challenge is semantic gap between the user defined rich semantics and the system defined low level features of the scene. In the present article, we propose an algorithm of CBVR using dynamic textures. Statistical textures with the inclusion of motion of the pattern, change in illumination of the pattern, intrinsic change to the pattern becomes dynamic textures. Dynamic textures are best suited to videos containing moving objects. In this article, LBP-TOP a variant of dynamic texture has been used as a feature for video retrieval. LBP-TOP has the capability of jointly describing motion and appearance features. The LBP-TOP features are invariant to illumination, rotation and local translation. These significant benefits support the use of LBP-TOP in the proposed method. The proposed method uses query video clip, which consists randomly selected ten example frames. The proposed CBVR has three stages offline processing, online processing, a matching & retrieval stage. In offline processing, we extract keyframes of database videos using Pearson Correlation Coefficient (PCC) and Color Moments (CM) and then LBP-TOP feature of keyframes have been extracted and used to represent entire database video. In online processing, we extract LBP-TOP features of query video and then these features will be given as input to matching & retrieval stage where, we calculate euclidean distance between LBP-TOP features of database keyframes and query video frames to retrieve videos with less distance. To prove effectiveness of the proposed method it have been tested on 108 videos of standard traffic dataset which is available publicly and compared with the other state-of-the-art methods, both qualitatively and quantitavely. Quantitative performance evaluation has been carried out using the evaluation parameters: Precision, Recall, Jaccard Index, Accuracy, Specificity and E-measure. Both qualitative and quantitative performance show that the proposed method performed well than the other state-of-the-art methods, and success of the proposed method lies under incorporation of dynamic textures. The proposed algorithm can be used in real-time applications like traffic monitoring. The proposed CBVR system can be used to monitor traffic through feature matching between query scene and the database video. If the query is matched with low traffic video of database then the algorithm displays output as low traffic time. In similar manner, medium traffic or heavy traffic times will be detected by the algorithm.

Keyword:

Pearson correlation coefficient (PCC)

Color moments

Keyframes

Dynamic textures

LBP-TOP

CBVR

Content based video retrieval using dynamic textures

摘要

期刊

机构

引用论文