![Galaxy Pattern
for instance
representation
Source: fig 1, from “Binary Descriptor Based Nonparametric Background Modelling for Foreground Extraction
by Using Detection Theory,”[1]](https://image.slidesharecdn.com/btpendsemppt-220218231559/85/Moving-object-detection-in-complex-scene-11-320.jpg)
![Galaxy Pattern
for instance
representation
Source: fig 1, from “Binary Descriptor Based Nonparametric Background Modelling for Foreground Extraction
by Using Detection Theory,”[1]](https://image.slidesharecdn.com/btpendsemppt-220218231559/85/Moving-object-detection-in-complex-scene-16-320.jpg)
![REFERENCES
[1] Min-Hsiang Yang, Chun-Rong Huang, Member, “Binary Descriptor
Based Nonparametric Background Modeling for Foreground Extraction by
Using Detection Theory,” IEEE Trans. on circuits and systems for video
tech., vol.25, no. 4, April 2015.
[2] Michael Calonder, Vincent Lepetit, O. Mustafa, T. Trzcinski, Christoph
Strecha and Pascal Fua, “BRIEF: Binary Robust Independent Elementary
Features, Computing a local binary descriptor very fast,” IEEE Trans.
Pattern Anal. Mach. Intell., vol.34, no. 7, July 2012.
[3] G.-H. Huang and C.-R. Huang, “Binary invariant cross color descriptor
using galaxy sampling,” in Proc. 21st Int. Conf. Pattern Recognit., Nov.
2012, pp. 2610–2613.
[4] Sample videos for testing source: CDnet 2012 benchmark data set and
CDnet 2014 benchmark data set.](https://image.slidesharecdn.com/btpendsemppt-220218231559/85/Moving-object-detection-in-complex-scene-31-320.jpg)
Moving object detection in complex scene
- 1. MOVING OBJECT DETECTION IN COMPLEX SCENES B. TECH. PROJECT PRESENTATION CPP 301 – CORE PROJECT I SUPERVISED BY: DR. S. MURALA
- 2. TEAM MEMBERS Abhishek Nandurkar Kumar Mayank Mayur Yadav Dept. of Electrical Engineering IIT Ropar Batch of 2017
- 3. ABSTRACT • Most background modelling approaches represent distribution of background changes by parametric models like Gaussian Mixture Models (GMM). • But complex variations of background changes are hard to be modelled by parametric background models. • Our work uses COARSE – TO – FINE DETECTION algorithm to extract foreground objects. • Model is updated by FIRST – IN – FIRST – OUT strategy to maintain the most recent observed background and foreground instance.
- 5. LITERATURE REVIEW • Parametric models like Gaussian Mixture Model and Bayesian learning-based model were mostly used. • These algorithms were time consuming. • A non-parametric model records pixels as samples of backgrounds and uses them as background model. • These showed superior performance than many other state-of-art methods. • Three categories of background modelling includes pixel- based, region based and hybrid methods.
- 6. MOTIVATION
- 7. “ • Background modelling is one of the important research topics for detecting moving foreground objects in visual surveillance. • Propose an efficient algorithm robust to illumination changes, noise, and dynamic background with minimal computational complexity. MOTIVATION
- 8. OUR PROCESS OF DETECTION OF MOVING OBJECTS BACKGROUND AND FOREGROUND MODEL CONSTRUCTION COARSE-TO-FINE DETECTION THEORY FOR FOREGROUND OBJECT EXTRACTION MODEL UPDATING
- 9. BACKGROUND AND FOREGROUND MODEL CONSTRUCTION • Background Model The background model is the union of individual background instances obtained from galaxy descriptor. • Foreground Model The foreground model is the union of individual foreground instances obtained from galaxy descriptor.
- 10. BACKGROUND AND FOREGROUND MODEL CONSTRUCTION Existing Algorithm • This model uses binary descriptors to extract foregrounds. • The descriptor used is called GALAXY. • Similar to actual galaxy the descriptor contains: o Galaxy center o Stars o Planets • Illumination changes are reduced by using transformed RGB color space.
- 11. Galaxy Pattern for instance representation Source: fig 1, from “Binary Descriptor Based Nonparametric Background Modelling for Foreground Extraction by Using Detection Theory,”[1]
- 12. OUR PROCESS OF DETECTION OF MOVING OBJECTS BACKGROUND AND FOREGROUND MODEL CONSTRUCTION COARSE-TO-FINE DETECTION THEORY FOR FOREGROUND OBJECT EXTRACTION MODEL UPDATING
- 13. DRAWBACKS • The above background modeling algorithm seems to be less efficient in few cases. • The shape of foreground regions we get in output are not precise. Also in case of small object, it fails to identify. • This method only compares relation between different colour intensities at a particular location. • The Spatio-Temporal Relations are not considered in this algorithm.
- 14. OUR PROCESS OF DETECTION OF MOVING OBJECTS BACKGROUND AND FOREGROUND MODEL CONSTRUCTION COARSE-TO-FINE DETECTION THEORY FOR FOREGROUND OBJECT EXTRACTION MODEL UPDATING
- 15. Our Proposed Algorithm • To create an instance, combines result of texture (gray scale channel) features and color space. • Uses same galaxy descriptor but creating an instance uses a whole different approach. • Considers temporal relations between two frames say, instance at ‘t’ will be combination of ‘t’ and ‘t-1’.
- 16. Galaxy Pattern for instance representation Source: fig 1, from “Binary Descriptor Based Nonparametric Background Modelling for Foreground Extraction by Using Detection Theory,”[1]
- 17. OUR PROCESS OF DETECTION OF MOVING OBJECTS BACKGROUND AND FOREGROUND MODEL CONSTRUCTION COARSE-TO-FINE DETECTION THEORY FOR FOREGROUND OBJECT EXTRACTION MODEL UPDATING
- 18. COARSE-TO-FINE DETECTION THEORY FOR FOREGROUND OBJECT EXTRACTION 1. COARSE-LEVEL DETECTION THEORY • An instance belongs to either a background or a foreground. • This forms a BINARY hypothesis. H0 : Ip(t’) is a background instance. H1 : Ip(t’) is a foreground instance. • Using MAP detection theory …… (1)
- 19. COARSE LEVEL DETECTION THEORY FOR FOREGROUND OBJECT EXTRACTION • Applying Bayes' rule, on equation (1) (2) • Hence, the likelihood ratio can be defined as (3) • For improving the computational complexity, we define (4) where dpB is the hamming distance
- 20. COARSE-TO-FINE DETECTION THEORY FOR FOREGROUND OBJECT EXTRACTION 2. FINE-LEVEL DETECTION THEORY • A pixel will belong to either foreground or background. • Considers a window of pixels, here 11x11. • A pixel is part of foreground or background. Also may/may not be part of above window. • Likelihood function is : (5)
- 21. OUR PROCESS OF DETECTION OF MOVING OBJECTS BACKGROUND AND FOREGROUND MODEL CONSTRUCTION COARSE-TO-FINE DETECTION THEORY FOR FOREGROUND OBJECT EXTRACTION MODEL UPDATING
- 22. MODEL UPDATING • Backgrounds will change with time, update of background models is an important issue. • Background models must contain most recent changes. • To ensure this, we have used: FIRST-IN-FIRST-OUT strategy to update background model.
- 24. Video I Sample video source: CDnet 2012 benchmark data set and CDnet 2014 benchmark data set.
- 25. Video I Sample video source: CDnet 2012 benchmark data set and CDnet 2014 benchmark data set.
- 26. Video I Sample video source: CDnet 2012 benchmark data set and CDnet 2014 benchmark data set.
- 27. Video I Sample video source: CDnet 2012 benchmark data set and CDnet 2014 benchmark data set.
- 28. APPLICATIONS • INTELLIGENT SURVEILLANCE FOR SECURITY • Border security • Company/Industry security • Home security • PARKING LOT MONITORING To have a count of the number of vehicles entering and exiting the lot.
- 29. FUTURE PLANS
- 30. FUTURE PLANS • Achieve better efficiency of the new model and get a better shape of the object. • We will Combine implemented texture based algorithm with color features along with spatio-temporal features for foreground extraction. • A new descriptor to represent the observed image regions. • We aim to develop a highly efficient algorithm robust to illumination effects and dynamic background. • Detecting a near perfect shape of the object is still a challenge.
- 31. REFERENCES [1] Min-Hsiang Yang, Chun-Rong Huang, Member, “Binary Descriptor Based Nonparametric Background Modeling for Foreground Extraction by Using Detection Theory,” IEEE Trans. on circuits and systems for video tech., vol.25, no. 4, April 2015. [2] Michael Calonder, Vincent Lepetit, O. Mustafa, T. Trzcinski, Christoph Strecha and Pascal Fua, “BRIEF: Binary Robust Independent Elementary Features, Computing a local binary descriptor very fast,” IEEE Trans. Pattern Anal. Mach. Intell., vol.34, no. 7, July 2012. [3] G.-H. Huang and C.-R. Huang, “Binary invariant cross color descriptor using galaxy sampling,” in Proc. 21st Int. Conf. Pattern Recognit., Nov. 2012, pp. 2610–2613. [4] Sample videos for testing source: CDnet 2012 benchmark data set and CDnet 2014 benchmark data set.
- 32. THANKS!
Editor's Notes
- #4: 2. significant illumination changes and dynamic moving backgrounds with time, 3. on basis of non parametric background and foreground models represented by binary descriptors.
- #5: Lets see the findings obtained from literature review.
- #6: Till now, various background modelling approaches have been developed, which can be categorized into pixel-based, region-based, and hybrid methods. One of the most famous pixel-based methods is Gaussian mixture models. A recursive Bayesian learning-based method was proposed next to detect foreground objects under dynamic backgrounds. They apply multilayer Gaussian distributions to model different background content. However, these algorithm are time consuming. A non-parametric model named as ‘Vibe’ was also proposed. It records pixels as samples of backgrounds in the surveillance video and uses them as background models instead of using an explicit parametric model for each pixel. It shows superior performance than other methods because it represents exact background changes in recent frames. However, they are easily affected by noise, illumination changes, and dynamic backgrounds. Hybrid methods, which integrate both pixel- and region-based methods, are best to resolve aforementioned problems. They can achieve better background representation but the computational complexity is relatively high.
- #7: So what motivated us to take this as our B.Tech Project.
- #8: We studied few of the background modelling and object detection methods mentioned above. We found that these methods have some drawbacks. Due to these difficulties, we decided to construct a model which can overcome above drawbacks. We aim to propose an efficient algorithm robust to illumination changes, noise, and dynamic background with minimal computational complexity.
- #23: Next step of our algorithm is background model updating. As backgrounds change with time, the update of background models is also an important issue to ensure that background models contain the most recent changes. Same goes with the foreground. We use first in first out strategy to update these models.
- #24: Lets see the results obtained by implementing the above algorithm discussed.
- #25: This is another video on which we tested our implemented algorithm. Based on the white patches seen in the detected output our algorithm clearly tracks the movement of these fishes. However it fails to detect the small fishes and sometimes unnecessary white patches appear. We will eradicate these problems in near future.
- #26: We tested our algorithm on 2 sample videos downloaded from CDnet 2012 benchmark data set and CDnet 2014 benchmark data set. It is a standard site from which videos are downloaded for testing. Left side of this detected video is the original video and right side is the detected output. White patches indicate foreground while black region represents background.
- #27: This is another video on which we tested our implemented algorithm. Based on the white patches seen in the detected output our algorithm clearly tracks the movement of these fishes. However it fails to detect the small fishes and sometimes unnecessary white patches appear. We will eradicate these problems in near future.
- #28: This is another video on which we tested our implemented algorithm. Based on the white patches seen in the detected output our algorithm clearly tracks the movement of these fishes. However it fails to detect the small fishes and sometimes unnecessary white patches appear. We will eradicate these problems in near future.
- #29: APPLICATION Lets see the real time application of our project. THE first one is Intelligent Surveillance for security In Border Security it can be used to warn us from possible threats at border from neighbouring countries. It can be also used in company/industry security – To keep a track of the people entering the office and prevent burglaries or accidents inside company/industry. Second application is Parking lot monitoring where it can be used to track the number of vehicles entering and exiting the lot.
- #30: So what motivated us to take this as our B.Tech Project.