A Survey on Various Methods for Animal Classification System

A Survey on Various Methods for Animal Classification System

Rakesh T M Computer Science REVA University

Bengaluru, Karnataka, India

Vishwanath R Hulipalled Computer Science REVA University

Bengaluru, Karnataka, India

AbstractAnimal classification based researches are advantageous for many real life applications. Animal classification systems are beneficial on the research connected to train behavioral of focused animal and also harmful animal disturbance in domestic area can be avoided. Studies based on animal classification plays a very vital role in quite a few real life applications. Applications which are very significant using this domain are avoiding animal vehicle collision on national highways, stopping harmful animal invasion in residential area, having knowledge on behavioral of targeted animal and many more. There are incomplete areas of research related to animal classification. Resourceful and dependable nursing of wild animals in their natural environments is crucial to inform safeguarding and management decisions. Advanced cameras are being a more and more popular tool for wildlife nursing due to their efficiency and dependability in gathering data of animals in wildlife discreetly. Nevertheless, working on such a huge bulk of images and videos captured from advanced camera manually is really difficult. This leads to the major problem to scientists and researchers to classify animals. In this paper we have done a rigorous survey on some of these areas for classification of animals.

KeywordsAnimal classification; Threshold; Segmentation Digital image processing; natural scene

1. INTRODUCTION

   Animal classification is a very significant and upcoming area due to a huge number of real life applications. Numerous animal classification methods and cautionary systems are used for alerting the presence of animals on the highways or residential area. All these applications can be divided into three areas namely classification, tracking and identification of animals. The first phase that is classification of animals is used in various meadows of real life applications. As an instance thousands of animal-vehicle mishaps were reported every day causing frequent deaths and vehicle damage. Examines regarding classification of animals in image processing have been an important field.

   Various algorithms and methods have been advanced by humanoid being in order to have a superior knowledge on animal behavior. Further, these applications also can be used as an alerting system to humanoid being from disturbance of hazardous wild animal for early protection actions. The first phase, which is the animal classification, has been applied in various fields of real life application. As an example, a classification algorithm has been advanced for light classification and ranging data to acknowledge fisherman to find the exact place of fishes in deep water.

   The second phase, which is the animal tracking, is the major subject in nursing animal behavior and its interface with the atmosphere. Technologies like sensor, radio-frequency identification, and global positioning system are used to track animals, animal tracking applications are trending in development of new tracking systems for animal identification of animal in forests or in highways.

   The third phase, which is the animal identification, will be helpful to identify the recognized animal. However, there are a lot of keys, ill-treatment of animals and threats in animal health are cumulative. To superior be able to the animals in vigorous information retrieving, location tracking, and RFID-based mobile nursing system has been intended to help users over a wireless network. Identification of animal has helped humanoid being to monitor and manage their animals easier and also helped in avoiding animal-vehicle collisions in highways. This paper will give more consideration and review for animal classification methods. The survey also narrows to the methods that use digital images or digital video. This survey will be given in the next section.

2. LITERATURE SURVEY

   O. Chapelle et al., [1] have worked on Identification of animal has aided humanoid being to observe and to take care their animals easily. Their work shows that support vector machines (SVMs) can simplify well on problematic image classification problems where the solitary features are high dimensional histograms. Their work also trails an experimental method, and its group clarifies as progressively better results are gotten over alterations of the SVM architecture and they have explained that it is likely to push the classification presentation obtained on image histograms to high levels with error rates as low as 11% for the classification of 14 Corel groups and 16% for a more general set of substances. Higher- level spatial features, histogram features are used. Histograms are used to bifurcate other types of data than images.

   N. Haering et al., [2] have proposed three-level video- event recognition methodology and applied it to animal-hunt discovery in wildlife biopics. In the first level they extracted color, texture, and motion topographies, and detected shot limits and moving object blobs. The mid-level works on neural network to find the object class of the moving object blobs. In this level they also generated shot descriptors that merges feature from the first level and implications from the mid-level. The shot descriptors are then used by the domain-specific implication procedure at the third level they detected video

   segments that match the user-defined happening model. The proposed method has been applied to the discovery of searches in wildlife documentaries. Event-based video indexing, summarization, and browsing are among the applications of their proposed method. In their work they have discussed on computational method and a number of enabling algorithmic components for automatic event discovery in video and applied it to detect searches in wildlife documentaries.

   Deva Ramanan et al., [3] have proposed object classification system and claimed it is capable of accurately detecting, localizing, and recovering the kinematic configuration of textured animals in real images. They have built a deformation model of shape from videos of animals and an appearance model of texture from a labeled group of animal images, and syndicate the two models spontaneously. They developed a simple texture descriptor. They tested their animal models on two datasets; images taken by professional photographers from the Corel collection, and mixed images from the web resumed by Google. Rather than using a histogram of textures, they represented texture with a patch of pixels. They demonstrated that their work is good for detecting animals. Broadly speaking, they also discussed on unsupervised system for learning articulated models using both video and images. These learned models appear promising for classification tasks beyond discovery, such as localization, kinematic recovery, and counting.

   S. L. Hannuna et al., [4] have described a novel method to detect mobile quadrupeds in complete wildlife film. Adjustable lighting, moving circumstances and inconspicuous animals make outmoded foreground extraction techniques such as optical flow and background subtraction unstable. They tracked a sparse set of points on short film clip and intercalate dense flow, using normalized convolution. Principal component analysis (PCA) is applied to a set of dense flows, describing animal gait and other activities. The forecast constants for significant principal components are examined as one dimensional time series. Projectin coefficient variation reproduces in the velocity and relative arrangement of the components of the foreground object. These constants relative phase changes are used to train a KNN classifier which segments the training data with 93% success rate. By making projection coefficients for hidden footage, their system has successfully located examples of animal gait beforehand missed by humanoid spectators.

   D. Ramanan et al., [5] contended on their work that tracking, object discovery, and model building are all alike activities. They described an automatic system that builds 2D spoken models known as pictorial structures from videos of animals. The system can be watched as a widespread tracker. The learned model can be coordinated to a visual library; here, the scheme can be viewed as a video classification algorithm. The learned model can also be recycled to notice the animal in imagesin this case, the scheme can be seen as a technique for learning models for object classification. They found that they can meaningfully improve the pictorial structures by supplementing them with a discriminative texture model learned from a texture library. They advanced a novel texture descriptor that outperforms the state-of-the-art for animal textures. They also have demonstrated the entire system on real video sequences of three dissimilar animals. They showed that

   they can track and identify the given animal. They also have used the learned models to identify animals from two data sets; comparing their results with simple baselines, they showed that, for the Google set, they can detect, localize, and recover part articulations from a collection demonstrably hard for object classification.

   J. C. Nascimento et al., [6] projected innovative methods to estimate the performance of object discovery algorithms in video sequences. The mentioned procedure highlights characteristics. Their proposed work matches the output of the algorithm with the ground truth and procedures the changes conferring to objective metrics. They claim it is possible to compare dissimilar methods, evaluating the métiers and feebleness and allowing the user to work on a reliable choice of the best algorithm for a specific application. They have applied this methodology to segment. These methods were calculated to assess how well it is possible to detect moving objects in an outside scene in fixed-camera.

   D. Duran et al., [7] proposed a new tiered heterogeneous wireless image sensor network for real-time inconspicuous species discovery and video classification based on a new Hierarchically Scalarized Character Oriented Discovery (HIS-CODE) algorithm and architecture. A tested test bed has been set up at the Omaha Henry Doorly Zoo, and data composed over wireless network from this test bed has been managed and classified based on the proposed HIS-CODE. This work backs to a novel combination in wireless sensor networks with the in-network scattered image classification functionality, particularly with the contemplation of natural world species image classifications.

   B. Kitt et al., [8] worked on new method for dynamic scene insight from a moving vehicle armed with a stereo camera. The proposed method is exclusively based on graphic evidence; hence it is appropriate to a large class of independent robots employed in interior as well as in open-air surroundings. Their projected method includes a motion estimation founded on disparity and optical flow using the Longuet-Higgins- Equations combined with an implicit extended Kalman-Filter. Based on this motion approximation a stirring object discovery and tracking is achieved. Each detected object is characterized with a unique ID while noticeable in the images. The projected algorithm was assessed on many stimulating real world image classifications.

   A. Kyme et al., [9] have worked on Non-invasive functional imaging of wide-awake, uncontrolled minor animals by means of motion-compensation eliminates the need for sedatives and allows an animals behavior response to incentives or managed drugs to be studied parallel with imaging. While the feasibility of motion-compensated radio tracer imaging of awake rodents using marker-based optical motion tracking has been exposed, marker less gesture tracking would avoid the danger of marker detachment, streamline the experimental workflow, and possibly provide more accurate pose estimates over a greater range of motion. Marker less motion tracking also led to a considerable lessening in motion artifacts in motion-compensated positron emission tomography imaging of a live, un-anesthetized rat. The results propose that further developments in live subjects are likely if non-rigid features are distinguished robustly and excepted from the pose approximation process.

   Z. Cao et al., [10] have absorbed on Online robotic labeling of oceanic animals in such video clips includes of three main steps: discovery and tracking, feature extraction and classification. The last two features are the focus of this work. Feature mined from convolutional neural network (CNN) is verified on two real-world oceanic animal datasets and produces improved classification consequences than current

3. GENERAL STRUCTURE AND METHODOLOGY FOR ANIMAL CLASSIFICATION SYSTEM IN

   EXISTING METHODS

   Training Phase Testing Phase

   methods. Suitable mixture of CNN and hand-designed features can attain even advanced accuracy than applying CNN alone. The group feature assortment scheme, which is an adapted version of the minimal-redundancy-maximal-relevance (mRMR) algorithm, helps as the standard for choosing an optimal set of hand-designed features. Performance of CNN and hand-designed features are additional inspected for images with dropped quality that outdoes bad lighting condition in water.

   V. Gan et al., [11] have discussed on nursing nocturnal giraffe behavior by plummeting numerous hours of thermal camera investigation footage into a short video swift which can be studied by humanoid forecasters. They have expressed the video summarization task as a tracking problem: frames in which giraffes are tracked are supposed to be characteristic poses/behaviors and not comprised in the swift, whereas frames

   Image Acquisition

   Pre-

   Animal Discovery

   Animal Classification

   Image Acquisition

   Pre- processing

   Animal Discovery

   where tracks reset or dismiss are presumed to be atypical events and are therefore included in the summary. To tool their tracking-by-discovery summarization method, they explored various mixtures of image features for long wave infrared spectrum cameras, and devise a parts-based object discovery method using geodesic detachments to handle the extreme variations of typical giraffe postures. Finally, they evaluated their summarization presentation in terms of recall and compressibility, and show how a trade-off exists between these two events using more fragile or robust tracking methods.

   S. C. Tan et al., [12] have presented that a supervised and unsupervised founded classification system to categorize the animals. Initially, the animal images are segmented using maximal region amalgamation segmentation algorithm. The Gabor features are removed from segmented images. Further, the removed features are reduced based on supervised and unsupervised methods. In supervised method, they have used Linear Discriminate Analysis (LDA) dimension discount technique to decrease the features. The decreased features are fed into symbolic classifier for the drive of classification. In unsupervised method, they have used Principle component analysis (PCA) dimension reduction method to decrease the features. The decreased features are fed into K-means algorithm for the purpose of grouping. Experimentation has been showed on a dataset of more than 2000 animal images containing of 20 dissimilar categories of animals with variable fractions of training samples. From the projected model, it is detected that supervised classification system achieves better associated to unsupervised method.

   Images

   As shown in figure 1 Animal classification is done in two phases. The very first phase is training phase followed by second phase i.e., testing phase. In the training phase images will be acquired either by capturing from still camera or by coral dataset available in the internet databases. These raw images wont be suitable for extracting the information as it consists of noises color distortion etc. In order to overcome from this preprocessing of the images is very much necessary. In the preprocessing domain images will be subjected to remove unwanted data like noise, light effects and image size formations. After cleaning the images are subjected to detect the objects in the images using suitable algorithms. Same techniques will be applied in the testing phase also. Both the images will be compared with the features to classify the detected animals in the images. Classification will be done based on error rates approaches. False acceptance rate and false rejection rate. The two falsely accepted rates average is taken called as estimated error rate. Error rate calculation based approach gives more accurate results than just ordinary classified data.

4. VARIOUS ANIMAL DISCOVERY METHODS IN IMAGE AND VIDEO PROCESSING

   Researches based on animal discovery plays a very vital role in many real life applications. Solicitations which are very important are stopping animal vehicle collision on roads, stopping unsafe animal intrusion in residential area, knowing train behavioral of targeted animal and many more.

   1. Humanoid Prediction Method for Animal Discovery

      Early investigates on animal discovery are based on to observe how fast and accurate humanoid eyes can detect the presence of animals in original image. This method is very good and reliable if the animal discovery distance is

      near and doesnt have lighting problems. This method for animal discovery by humanoid eyes is also reliable if seen from the computational point of view. Work done in [6] presented that a humanoid observer is able to make a decision whether a briefly flashed animal image is consuming the presence of an animal as wicked as 150ms. Even though this method of humanoid prediction for animal discovery is effective and achieves some reasonable result or level, humanoid eyes do have some serious limitations. Humanoid eyes can get tired or exhausted easily producing a curb in the effectiveness and accuracy of the method (algorithm). Humanoid eyes need some break and cant work efficiently for whole day to accomplish animal discovery. These limitations can be restricted by using computer vision in image processing for animal discovery [13].

   2. Threshold based Segmentation Method for Animal Classification

      To remove the beleaguered animals RoI from background, this approach can be secondhand. The knowledge of this method is modest in which the pixels in the image having strengths or values greater than the threshold are usual to white and those pixels having strengths or values less than the threshold value are set to black. There are dissimilar types of thresholding like adaptive thresholding or dynamic thresholding and best thresholding which are very significant topics image processing but in this work they have limited to modest concept of thresholding only the object or animal is originate by using background subtraction method after getting the background image. Work done in [8] shows that it is very problematic and deadly to choose the threshold value as the background image changes periodically [14].

   3. Control Range based Method for Animal Discovery

      Researchers have strained to find out whether the company of animal in the scene or image will affect the control range of the image or not which can be defined as the amplitude of the signal in the frequency domain. The control range can be built by altering images from area to frequency domain with the help of the alteration function like Fourier transform. Work carried out in [15] shows that this method is not suitable if a person wants quick result or needs to notice the animals very quickly as this technique takes more time.

   4. Face Unearthing based Method for Animal Discovery

   To monitor or observe the behavior of animals and their communication with the environs, Z. Cao et al., [10] applied discovery and tracking of animal faces using Haar-like feature

   and Adaboost classifiers. When it is optimistic that animal has been identified, video plotters twisted on to spread to ensure that logged video covers a precise investigation value. This approach is very vital and significant in state whereby person is not appropriate to current at the footage scene for care subject or person might be frightened off some timorous animal away. The dimension of animal faces is done by using face detection method with dissimilar local contrast configuration of luminescence channel to detect the image region of animal faces.

5. CONCLUSION

   There are a lot of glitches essential to be measured in emerging an animal classification algorithm. Primary is the illumination problem, in which an unexpected change of illumination result typically in interior request can affect the efficiency in noticing the attendance of animal interruption. Also, luminance problematic with vicissitudes of usual setting from day to night at outside shadowing system can also touch the detection. Also, affecting background, such as greeneries by breeze strength be stared as forefront image and some sedentary animal which continue still for an extended time can be incorrectly understood as contextual image by the algorithms. There are lots of problems that essential to be taken into account for developing an efficient animal discovery algorithm. Animals originate from nowhere so we cant forecast their occurrence and also the rapidity of the animals cant be checked or noticed. There is the illumination problem also, where an unexpected alteration of illumination result can affect the efficacy in detecting the presence of animal intrusion. Also each animal has its own physiognomies and conduct with the environs which leads to a problem in identification of correct animals. So in forthcoming heaps of work and investigates needs to be done in emerging a real animal discovery algorithm.

   Table 1 provides the summary of various methods used for animal classification system. From the above literature survey that is discussed so far and also from the comparison analysis of dissimilar methods provided in table 2, it is evident that CNN method is the best algorithm for animal classification system.

6. FUTURE SCOPE

   As identified in this survey, convolutional neural network is the best approach for animal classification. As a future scope one can take low-resolution datasets, the image size with 32 x 32 pixels and plan to train the model for 6 to 8 categories. The advantage of training the model with low-resolution images is

   – it will be useful in the military to classify the objects in long- distance. Usually in long-distance images, the quality of the images are with low quality. So if one could classify the object with the low-resolution, then it will be useful to analyze long- distance captured images.

   TABLE 1: ANALYSIS OF VARIOUS METHODS FOR ANIMAL CLASSIFICATION SYSTEM

   Sl. No.	Title	Methodology Identified	Results identified in terms of accuracy	Limitations and Recommendations
   1	Support Vector Machines for Histogram-Based Image Classification [1].	Support Vector Machine, Image Histograms	Using Image Histograms error rate found about 11% & 16% for generic set of objects	Limitations: accuracy varies for dissimilar types of inputs. No constant output using the combinations of proposed method. Recommendations:we can use multimodal methods to get accuracy results for all type of datasets.
   2	A Semantic Event-Discovery Method and Its Application to Detecting Hunts in Wildlife Video [2].	Video-event- discovery, Event based video indexing	Actual Frames: approximately I, 00,000. Detected-Frames: approximately 95,000. Precession Rate: 95%, Recall Rate: 86%	Limitations: when frames are less accuracy is more but higher the frames lesser the accuracy is identified in the proposed method. Recommendations: accuracy should be constant for even for the higher frames.
   3	Detecting, Localizing and Recovering Kinematics of Textured Animals [3].	Shape & Texture extraction models using SIFT descriptors	For Corel data set:85% For google Data set: 94%	Limitations: accuracy varies for dissimilar types of inputs. No constant output using the combinations of proposed method. Recommendations: we can use multimodal methods to get accuracy results for all type of datasets.
   4	Identifying Quadruped Gait In Wildlife Video [4].	Principal Component Analysis using KNN Classifier	93% accuracy for high quality images.	Limitations: Few gaits misclassified and two examples of non-gait are classified as gait. Worked on only limited trained data sets Recommendations: Results should give more accuracy even with huge datasets.
   5	Building Models of Animals from Video [5].	Tracking, video analysis, object recognition, texture, shape	For Corel data set:77% For google Data set: 68%	Limitations: accuracy varies for dissimilar types of inputs. No constant output using the combinations of proposed method. Recommendations: we can use multimodal methods to get accuracy results for all type of datasets.
   6	Performance Evaluation of Object Discovery Algorithms for Video Surveillance [6].	Ground truth, metrics, multiple interpretations, performance evaluation, segmentation, surveillance systems.	Correct Discovery using SGM Algorithm: 91.2%. Discovery failure: 8.5%. Matching area 78.8%	Limitations: accuracy varies for correct discovery and failed discovery using the combinations of proposed method. Recommendations: Matching area should be 90% above accuracy if we say accuracy discovery is above 90%
   7	Hierarchical Character Oriented Wildlife Species Recognition Through Heterogeneous Wireless Sensor [7].	Hierarchically Scalarized Character Oriented Discovery (HIS-CODE) algorithm	Deviation: 25.2% for bigger animals & 7.6% for smaller animals. Success rate 90%	Limitations: Deviation should be very less for any type given RoI. Recommendations: Should identify the algorithms which gains less deviations in recognizing the objects. Which increases the success rate to 100%
   8	Discovery and Tracking of Independently Moving Objects in Urban Environments [8].	Optical flow using the Longuet Higgins- Equations with an implicit extended Kalman-Filter	Results have been only discussed in terms of high and low in their work. % wise results have not been mentioned.	Limitations: Analytical results havent been discussed in any part of the work to analyze the mentioned work. Recommendations: Analytical or Graphical results gives more value
   9	Markerless Motion Tracking of Awake Animals in Positron Emission Tomography [9].	Markerless optical motion tracking, motion compensation, positron emission tomography (PET)	Local Contrast: 88%	Limitations: Results have been generated for only rat identification. Accuracy is also less compared to other algorithms. Recommendations: Results should be generated by taking examples of all types of animals.
   10	Marine Animal Classification Using Combined CNN and Hand-designed Image Features [10].	convolutional neural network (CNN)	Classification error of images with degraded quality.3.0+/-0.5	Limitations: Discussed method using only one method i.e., using CNN. Cant predict this is best

   					with any comparison results using dissimilar algorithms.
   	11	Monitoring Giraffe Behavior in Thermal Video [11].	Unnormalized gradient histograms, SVM, HoG	Multiframe tracking: 0.9 Twoframe tracking- 0.5	Limitations: Results varying from multi frame to two frames. Usually accuracy in tracking should be consistent in given any frames input.
   	12	Supervised and Unsupervised Learning in Animal Classification [12].	PCA, LDA, Symbolic Representation, Clustering	For 70% of training data using supervised 79.54% accuracy, For 70% of training data using unsupervised 75.46% accuracy	Limitations: Accuracy is too low compared to other existing techniques.
   	13	AnimalVehicle Collision Mitigation System for Automated Vehicles [13].	HoG, LBP-ADA boost, SVM	HAAR: Maean of TPR: .9 HAAR: Maean of FPR: .02 LBP: Maean of TPR: .9 LBP: Maean of FPR: .002 HoG: Maean of TPR: .9 HoG: Maean of FPR: .02	Limitations: Less number of features have been taken in every comparison analysis. More the number features more will be the accuracy.
   	14	A Practical Animal Discovery and Collision Avoidance System Using Computer Vision Technique [14]	Cascade classifier, computer vision, haar, image processing, OpenCV	Accuracy: 82.5%	Limitations: Accuracy is too low compared to other existing techniques.
   	15	Animal Recognition and Identification with Deep Convolutional Neural Networks for Automated Wildlife Monitoring	deep learning, convolutional neural networks	Detecting containing single animals: 96.6%. Detecting containing multiple animals: 96.6%.	Limitations: Accuracy is varying with number animals present in the images. Accuracy should be constant for any number of given animals in the given input image.

   TABLE 2: COMPARISON ANALYSIS OF DISSIMILAR METHODS

   Sl No	Methods	Approximate Accuracy
   1	Support Vector Machine	89%
   2	Event based video indexing	95%
   3	Shape & Texture extraction models using SIFT descriptors	85%
   4	Principal Component Analysis using KNN Classifier	93%
   5	Tracking, video analysis, object	77%

   Sl No	Methods	Approximate Accuracy
   6	SGM algorithm	91%
   7	HIS-Code Algorithm	90%
   8	Optical flow	90%
   9	Markerless optical motion tracking	88%
   10	convolutional neural network (CNN)	96%
   11	PCA, LA	80%
   12	Cascade classifier, computer vision	85%

   Sl No	Methods	Approximate Accuracy
   1	Support Vector Machine	89%
   2	Event based video indexing	95%
   3	Shape & Texture extraction models using SIFT descriptors	85%
   4	Principal Component Analysis using KNN Classifier	93%
   5	Tracking, video analysis, object	77%

   Sl No	Methods	Approximate Accuracy
   6	SGM algorithm	91%
   7	HIS-Code Algorithm	90%
   8	Optical flow	90%
   9	Markerless optical motion tracking	88%
   10	convolutional neural network (CNN)	96%
   11	PCA, LDA	80%
   12	Cascade classifier, computer vision	85%

   Graphical Analysis of dissimilar methodologies

   120

   100

   80

   60

   40

   20

   	Support Vector Machin e	Event based video indexing	Shape & Texture extracti on models using SIFT descri	Principa l Compon ent Analysis using KNN Classifi	Tracking , video analysis, object recognit ion, texture, shape	SGM algorith m	HIS- Code Algorith m	Optical flow	Markerl ess optical motion tracking ,	convolu tional neural network (CNN)	PCA, LDA,	Cascade classifie r, comput er vision,
   Column1	89	95	85	93	77	91	90	90	88	96	80	85

   	Support Vector Machin e	Event based video indexing	Shape & Texture extracti on models using SIFT descri	Principa l Compon ent Analysis using KNN Classifi	Tracking , video analysis, object recognit ion, texture, shape	SGM algorith m	HIS- Code Algorith m	Optical flow	Markerl ess optical motion tracking ,	convolu tional neural network (CNN)	PCA, LDA,	Cascade classifie r, comput er vision,
   Column1	89	95	85	93	77	91	90	90	88	96	80	85

   0

   89 95 85

   93 91 90 90 88 96

   77

   80 85

   Fig. 2. Graphical Analysis of Dissimilar Methodologies

   As Shown in Fig 2, graphical analysis has been done for the commonly and effectively used algorithms and also shows the result comparison of various methods in animal tacking, identifying and classification methods. Amongst all the methods it is identified that Convolution Neural Network (CNN) is best in giving the accurate results.

7. REFERENCES

1. O. Chapelle, P. Haffner and V. N. Vapnik, "Support vector machines for histogram-based image classification," in IEEE Transactions on Neural Networks, vol. 10, no. 5, pp. 1055-1064, Sept. 1999.

2. N. Haering, R. J. Qian and M. I. Sezan, "A semantic event-discovery method and its application to detecting hunts in wildlife video," in IEEE Transactions on Circuits and Systems for Video Technology, vol. 10, no. 6, pp. 857-868, Sept. 2000.

3. Deva Ramanan, D. A. Forsyth and K. Barnard, "Detecting, localizing and recovering kinematics of textured animals," 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), San Diego, CA, USA, 2005, pp. 635-642 vol. 2.

4. S. L. Hannuna, N. W. Campbell and D. P. Gibson, "Identifying quadruped gait in wildlife video," IEEE International Conference on Image Processing 2005, Genova, 2005, pp. I-713.

5. D. Ramanan, D. A. Forsyth and K. Barnard, "Building models of animals from video," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 28, no. 8, pp. 1319-1334, Aug. 2006.

6. J. C. Nascimento and J. S. Marques, "Performance evaluation of object discovery algorithms for video surveillance," in IEEE Transactions on Multimedia, vol. 8, no. 4, pp. 761-774, Aug. 2006.

7. D. Duran, D. Peng, H. Sharif, B. Chen and D. Armstrong, "Hierarchical Character Oriented Wildlife Species Recognition Through Heterogeneous Wireless Sensor Networks," 2007 IEEE 18th International Symposium on Personal, Indoor and Mobile Radio Communications, Athens, 2007, pp. 1-5.

8. B. Kitt, B. Ranft and H. Lategahn, "Discovery and tracking of independently moving objects in urban environments," 13th International IEEE Conference on Intelligent Transportation Systems, Funchal, 2010, pp. 1396-1401.

9. A. Kyme et al., "Markerless Motion Tracking of Awake Animals in Positron Emission Tomography," in IEEE Transactions on Medical Imaging, vol. 33, no. 11, pp. 2180-2190, Nov. 2014.

10. Z. Cao, J. C. Principe, B. Ouyang, F. Dalgleish and A. Vuorenkoski, "Marine animal classification using combined CNN and hand-designed image features," OCEANS 2015 – MTS/IEEE Washington, Washington, DC, 2015, pp. 1-6.

11. V. Gan, P. Carr and J. Soltis, "Monitoring Giraffe Behavior in Thermal Video," 2015 IEEE Winter Applications and Computer Vision Workshops, Waikoloa, HI, 2015, pp. 36-43.

12. S. C. Tan, "Using Supervised Attribute Selection for Unsupervised Learning," 2015 4th International Conference on Advanced Computer Science Applications and Technologies (ACSAT), Kuala Lumpur, 2015, pp. 198-201.

13. A. Mammeri, D. Zhou and A. Boukerche, "Animal-Vehicle Collision Mitigation System for Automated Vehicles," in IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 46, no. 9, pp. 1287-1299, Sept. 2016.

14. S. U. Sharma and D. J. Shah, "A Practical Animal Discovery and Collision Avoidance System Using Computer Vision Technique," in IEEE Access, vol. 5, pp. 347-358, 2017.

15. H. Nguyen et al., "Animal Recognition and Identification with Deep Convolutional Neural Networks for Automated Wildlife Monitoring," 2017 IEEE International Conference on Data Science and Advanced Analytics (DSAA), Tokyo, 2017, pp. 40-49.