Smart Spaces Lab

Computer Vision based Asset Surveillance for Smart Buildings

Research Team
Latha Parameswaran, Gautam, K. S.,and Senthil Kumar Thangavel.
Unraveling meaningful pattern form the video offers a solution to many real-world problems, especially surveillance and security. Detecting and tracking an object under the area of video surveillance, not only automates the security but also leverages smart nature of the buildings. The objective of the manuscript is to detect and track assets inside the building using vision system. In this manuscript, the strategies involved in asset detection and tracking are discussed with their pros and cons. In addition to it, a novel approach has been proposed that detects and tracks the object of interest across all the frames using correlation coefficient. The proposed approach is said to be significant since the user has an option to select the object of interest from any two frames in the video and correlation coefficient is calculated for the region of interest. Based on the arrived correlation coefficient the object of interest is tracked across the rest of the frames. Experimentation is carried out using the 10 videos acquired from IP camera inside the building.

A Cascade Color Image Retrieval Framework for Image Retrieval

Research Team
Latha Parameswaran, Gautam, K. S., Senthil Kumar Thangavel.
The search for the digital image from the repository is challenging since the volume of the image created and consumed is growing exponentially with respect to time. This makes the image retrieval an ongoing research problem. Rather than relying on metadata, analyzing the content of the image is proven to be a successful solution for retrieval. Since manually annotating the growing images is considered to be impossible. Thus the need for the hour is a framework that is capable of retrieving the similar images with less time complexity. In this paper, an image retrieval framework has been proposed to retrieve similar images from the repository. The motivation of the work is to leverage the smart nature of building, in such a way that when an asset inside the building is captured and given as query image, the user of the building should be provided with all the relevant assets inside the building. The proposed framework is built with color, shape and edge retrieval system that works in cascade approach. Since the framework works as a cascade, it is observed that the results get fine-tuned at every layer of the cascade thus increasing the precision. The novelty of the work lies in the third layer where the XOR operation is performed to check the magnitude of dissimilarity between the query and database images. Based on the above dissimilarity, the threshold has been fixed to differentiate the image of interest from other images in the repository. The performance of the proposed approach is evaluated with the manually built dataset. On evaluating the performance, it is inferred that the precision of the proposed framework is 99%.

A Computer Vision Based Approach for Object Recognition in Smart Buildings

Research Team
Kumar, Latha Parameswaran, D. Kavin, and Senthil Kumar Thangavel
Object recognition is one of the essential Computer Vision techniques. The success of object recognition lies in identifying features that strongly represent the object of interest. The manuscript comes up with a hybrid feature descriptor that combines the properties of HOG, ORB and BRISK feature descriptors. Linear SVM is used to classify the feature vectors of the object of interest and other objects in the scene. Occlusion, Orientation and Scaling are some of the limitations in existing approach. From the experimental analysis, we infer that the proposed framework handles partial occlusion and is invariant to scaling and rotation. The framework has been tested with a manually built data library and the classification accuracy of the proposed framework is 0.91, whereas the standalone performance of the HOG, ORB and BRISK are 0.85, 0.87, and 0.89 respectively.

An Efficient Rule Based Algorithm for Fire Detection on Real Time Videos.

Research Team
Latha Parameswaran,Sridhar, P., and Senthil Kumar Thangavel.
The projected work shows generic rule in YCbCr color space based fire pixel detection is proposed for smart building which will complement the conventional electronic sensor based fire detection system. The proposed method handles YCbCr color model is used for decoupling the luminance and chrominance which added discriminate the color than RGB color model. This algorithm has been tested on fire and fire like images which results in 97.95% detection accuracy. Obtained experimental results have been compared with other existing algorithms and it is observed that gives a very high the proposed algorithm detection accuracy and feasible true positive rate in fire images.

Video-based fire detection by transforming to optimal color space

Research Team
Sridhar, P, Manickam, M. T., Yogesh, M., Thangavel, S. K., & Parameswaran, L.
Unraveling meaningful pattern form the video offers a solution to many real-world problems, especially surveillance and security. Detecting and tracking an object under the area of video surveillance, not only automates the security but also leverages smart nature of the buildings. The objective of the manuscript is to detect and track assets inside the building using vision system. In this manuscript, the strategies involved in asset detection and tracking are discussed with their pros and cons. In addition to it, a novel approach has been proposed that detects and tracks the object of interest across all the frames using correlation coefficient. The proposed approach is said to be significant since the user has an option to select the object of interest from any two frames in the video and correlation coefficient is calculated for the region of interest. Based on the arrived correlation coefficient the object of interest is tracked across the rest of the frames. Experimentation is carried out using the 10 videos acquired from IP camera inside the building.

Dynamic and chromatic analysis for fire detection and alarm raising using real-time video analysis Advances in Intelligent Systems and Computing

Research Team
Sridhar, P, Srishilesh, P. S., Parameswaran, L., Tharagesh, R. S., Thangavel, S. K.
Fire outbreak has become a common accident that occurs in several places such as in forests, manufacturing industries, living house and in widely crowded areas. These incidents cause severe damage to nature as well as to living creatures in the affected surroundings. Due to this, the need for efficient fire detection system has been increased rapidly. Using fire detecting sensors has proved to be an efficient solution but its effectiveness on delivering quick results depends on the affinity of fire sources. In the proposed method, we present an economical and affordable fire detection algorithm using video processing techniques which is compatible with CCTV and other stationary surveillance cameras. The algorithm uses an RGB color model with chromatic and dynamic disorder analysis to detect the fire. Fire pixels are detected by the rules of the color model which is mainly dependent on the fire pixel intensity and also the saturation of red color component in the fire pixel. The extracted fire like pixels are authorized by growth combined with the disorder of the fire regions. Furthermore, based on iterative checking the real fire is identified, if it is present then the appropriate signals will be sent. The proposed method is tested on various datasets acquired in real time environments and from the internet. This methodology can be used for fully automatic fire detection surveillance with reduced false true errors.

Design of an Image Skeletonization Based Algorithm for Overcrowd Detection in Smart Building

Research Team
Latha Parameswaran, R. Manjusha
Crowd analysis has found its significance in varied applications from security purposes to commercial use. This proposed algorithm aims at contour extraction from skeleton of the foreground image for identifying and counting people and for providing crowd alert in the given scene. The proposed algorithm is also compared with other conventional algorithms like HoG with SVM classifier, Haar cascade and Morphological Operator. Experimental results show that the proposed method aids better crowd analysis than the other three algorithms on varied datasets with varied illumination and varied concentration of people.

Kernel Based Approaches for Context Based Image Annotatıon

Research Team
R. Manjusha,Swati Nair R. Manjusha, Latha Parameswara
The Exploration of contextual information is very important for any automatic image annotation system. In this work a method based on kernels and keyword propagation technique is proposed. Automatic annotation with a set of keywords for each image is carried out by learning the image semantics. The similarity between the images is calculated by Hellinger’s kernel and Radial Bias Function kernel(RBF)kernel. The images are labelled with multiple keywords using contextual keyword propagation. The results of using the two kernels on the set of features extracted are analysed. The annotation results obtained were validated based on confusion matrix and were found to have a good accuracy. The main advantage of this method is that it can propagate multiple keywords and no definite structure for the annotation keywords has to be considered

Detection and Removal of RainDrop from Images Using DeepLearning

Research Team
R. Manjusha, Y. Himabindu, Latha Parameswaran
Dynamic climatic conditions like rain, affects the performance of vision algorithms which are used for surveillance and analysis tasks. Removal of rain-drops is challenging for single image as the rain drops affect the entire image and makes it difficult to identify the background affected by the rain. Thus, removal of rain from still pictures is a complex and challenging task. The rain drops affect the visibility and clarity of the image which makes it difficult to read and analyze the information present in the image. In this paper, we identified and restored the rain drop affected regions using a deep learning architecture.

Vision Based Algorithm for People Counting using Deep Learning

Research Team
R. Manjusha, Padmashini, M., Dr. Latha Parameswaran
Estimating the number of people in a particular scene has always been an important topic of research in computer vision and digital image processing. People counting has wide applications in scenario ranging from analyzing the customer’s choice and improving the quality of service in retail stores, supermarkets and shopping malls to managing human resources and optimizing the energy usage in office buildings. While there exists algorithms for counting people in a scene, some algorithm have set their benchmark in performance with respect to efficiency, flexibility and accuracy. In this paper, an attempt has been made to perform people counting using Deep Neural Networks (DNN) on comparison with existing image processing based algorithms like Histogram of Oriented Gradients with Support Vector Machine (HoG with SVM), Local Binary Pattern (LBP) based Adaboost classifier and contour based people detection. The proposed DNN based approach has higher accuracy at 90% and less false negatives.

Funded Project Details

1. Srishilesh P S, Sanjay Tharagesh R S, P.Sridhar, Dr.Latha Parameswaran, Senthil Kumar Thangavel,” Dynamic and Chromatic Analysis for Fire Detection and Alarm Raising Using Real-Time Video Analysis“,Proceedings of 3rd International Conference On Computational Vision and Bio Inspired Computing,2019.
2. Kavin Kumar D, Latha Parameswaran,Senthil Kumar Thangavel,”A Computer Vision Based Approach for Object Recognition in Smart Buildings“,Proceedings of 2nd International Conference On Computational Vision and Bio Inspired Computing,2018.
3. Thanga Manickam M,Yogesh M, P.Sridhar ,Senthil Kumar Thangavel ,Latha Parameswaran,”Video-Based Fire Detection by Transforming to Optimal Color Space“,Proceedings of 3rd International Conference On Computational Vision and Bio Inspired Computing,2019.
4. Sridhar P,Latha Parameswaran,Senthil Kumar Thangavel,”An Efficient Rule Based Algorithm for Fire Detection on Real Time Videos“,Proceedings of the first international conference on Intelligent Computing,2018.
5. K S Gautam, Latha Parameswaran, Senthil Kumar Thangavel,“Computer Vision based Asset Surveillance for Smart Buildings“,Proceedings of the first international conference on Intelligent Computing,2018.
6. K S Gautam, Latha Parameswaran, Senthil Kumar Thangavel,“A Cascade Color Image Retrieval Framework for Image Retrieval“,Proceedings of 2nd International Conference On Computational Vision and Bio Inspired Computing,2018.
7. Harsh Motka, Latha Parameswaran, “A Vision Based Approach For Anomaly Detection In Smart Environments Using Thermal Images”, International Journal of Innovative Technology and Exploring Engineering (IJITEE), Volume-8 Issue-7, May 2019, Page No.2838-2844, ISSN: 2278–3075 (Online).
8. Sudipta Rudra,Senthil Kumar Thangavel,“A Robust Q-Learning and Differential Evolution Based Policy Framework for Key Frame Extraction“, Springer Advances in Intelligent Systems and Computing ,Vol.1039,pp.716-728,2019.
9. Padmashini, R. Manjusha, , Latha Parameswaran, “Vision Based Algorithm for People Counting using Deep Learning”, International Journal of Engineering and Technology”(UAE),Volume 7, issue 3, pp 74-80, 2018