Human Detection, Tracking and Trajectory Extraction in a

Human Detection, Tracking and Trajectory Extraction in a
International Journal of Modern Engineering Research (IJMER)
Vol. 3, Issue. 3, May - June 2013 pp-1843-1846
ISSN: 2249-6645
Human Detection, Tracking and Trajectory Extraction
in a Surveillance Camera network
Peyman Babaei
Dep. of Computer, West Tehran Branch, Islamic Azad University, Tehran, Iran
ABSTRACT: This paper proposes human tracking and recognition method in a camera network. Human matching in a
multi-camera surveillance system is a fundamental
issue for increasing the accuracy of recognition in multiple views of
cameras. In camera network, videos have different characteristics such as pose, scale and illumination. Therefore it is
necessary to use a hybrid scheme of scale invariant feature transform to detection
and recognition human’s behaviors. The
West to extract trajectories. Extracting the
main focus of this paper is to analyze activities
trajectories help to detect abnormal behavior which may be occluded in single- camera
North surveillance.
KEYWORDS: Camera network, Multi-camera
surveillance, Human’s behavior, Trajectories extraction.
I. 2nd
1st Qtr
Qtr 3rd Qtr 4th Qtr
Tracking and behavior recognition are two fundamental tasks in video surveillance systems which are widely
employed in commercial applications for purposes of statistics gathering and processing. The number of cameras and
complexity of surveillance systems have been continuously increasing to have better coverage and accuracy. Multi-camera
systems become increasingly attractive in machine vision. Applications include multi view object tracking, event detection,
occlusion handling and etc. In this paper, we develop method for tracking and recognition by a traffic video surveillance
system of two cameras with a partially overlapping field of view.
This paper is organized as follows: an overview of the past works in section2. Our proposed architecture and
algorithm is presented in section3. Results of subjective evaluations and objective performance measurements with respect to
Ground-truth are presented in section4. Section5 contains the conclusion.
In the last few years, a lot of works in detecting, describing and matching feature points has deployed. In a camera
network features’ matching between multiple images of a scene is an important component of many computer vision tasks.
Although the correspondences can be hand selected, such a procedure is hardly conceivable as the number of cameras
increases or when the camera configuration changes frequently, as in a network of pan-tilt-zoom cameras [1]. Other methods
for finding correspondences across cameras [2] have been developed through a feature detection method such as the Harris
corner detection method [3] or scale invariant feature transform [4]. In [5] shown that corners were efficient for tracking and
estimating structure from motion. A corner detector is robust to changes in rotation and intensity but is very sensitive to
changes in scale. The Harris detector finds points where the local image geometry has high curvature in the direction of both
maximal and minimal curvature, as provided by the eigen-values of the Hessian matrix. They develop an efficient method
for determining the relative magnitude of the eigen-values without explicitly computing them. Such color-based matching
methods have also been used to track moving objects across cameras [6, 7]. Scale invariant features matching were first
proposed in [8] and attracted the attention of the computer vision systems for invariant to scale, rotation, and view-point
variations. Also uses a scale-invariant detector in the difference of Gaussian (DOG) scale space. In [4] fits a quadratic to the
local scale-space neighborhood to improve accuracy. Then creates a Scale Invariant Feature Transform descriptor to match
key-points using a Euclidean distance metric in an efficient best-bin first algorithm where a match is rejected if the ratio of
the best and second best matches is greater than a threshold.
A comparative study of many local image descriptors [9] shows the superiority of this method with respect to other
feature descriptors for the case of several local transformations. In [10] develop a scale-invariant Harris detector that keeps
key points at each scale only if it’s a maximum in the Laplacian scale-space [11]. More recently, in [12] integrate edge-based
features with local feature-based recognition using a structure similar to shape contexts [13] for general object-class
recognition. In [14] propose a matching technique based on the Harris corner detector and a description based on the Fourier
transform to achieve invariance to rotation. Harris corners are also used in [15], where rotation invariance is obtained by a
hierarchal sampling that starts from the direction of the gradient. In [16] introduce the concept of maximally stable external
region to be used for robust matching. These regions are connected components of pixels which are brighter or darker than
pixels on the region’s contour; they are invariant to affine and perspective transform, and to monotonic transformation of
image intensities. Among the many recent works populating the literature on key-point detection, it is worth mentioning the
scale and affine invariant interesting points recently proposed in [17], as they appear to be among the most promising keypoint detectors to date. The detection algorithm can be sketched as follows: first Harris corners are detected at multiple
scales, and then points at which a local measure of variation is maximal over scale are selected. This provides a set of
distinctive points at the appropriate scale. Finally, an iterative algorithm modifies location, scale, and neighborhood of each
point and converges to affine invariant points. In [18] describe a matching procedure wherein motion trajectories of objects
tracked in different cameras are matched so that the overall ground plane can be aligned across cameras following a
homograph transformation [19-21].
1843 | Page
International Journal of Modern Engineering Research (IJMER)
Vol. 3, Issue. 3, May - June 2013 pp-1843-1846
ISSN: 2249-6645
First, we review the function of a typical single-camera and multi-camera surveillance system as presented in our
previous work [22], the function of a typical single-camera surveillance system is illustrated in Fig.1. The first part of the
processing flowchart is very general, which is marked “Detecting & Matching Features Extraction Pipeline”. This pipeline
may produce all target information (pose, scale, illumination, color, shape, etc.), and potentially the description of the scene.
The end of the processing pipeline, the human tracking and classification is done.
Detection& Matching Pipeline
Fig. 1: Single Camera Processing
Only the matching features have to be stored, instead of high quality video suitable for automated processing. This
method enables the multi-camera surveillance system. The video surveillance system, as described in the above, cannot
provide an adequate solution for many applications [23-27]. A multi-camera surveillance system tracking targets from one
camera to the next can overcome all these limitations. A typical multi-camera surveillance network is illustrated in Fig.2.
Fusing at the matching features level requires merging all the features from the cameras on to a full representation of the
environment. This approach distributes the most time consuming processing between the different cameras, and minimizes
communication, since only the extracted features needs to be transmitted, no video or image. Given these advantages, system
communicates only the matching features for fusion.
Fig. 2: Multi camera network Processing
The problem of multi-view activity recognition has been addressed in many papers, but almost the information of
multiple views is fused centrally. Our proposed framework is decentralized. The pose of cameras at intersection is shown in
Fig 3: Camera setup in a network
In Fig.4, the structure camera network is illustrated. Each of the cameras has processing cores in four levels. The
input stream is fed to detection level. At the decision level, control commands are issued to classify the detected human
based on extracted description features. Processing cores in three upper levels exchange the requisite information to track
and recognition more accurately.
Fig 4 Structure camera network
1844 | Page
International Journal of Modern Engineering Research (IJMER)
Vol. 3, Issue. 3, May - June 2013 pp-1843-1846
ISSN: 2249-6645
The Scale Invariant Feature Transform has been shown to perform better than other local descriptors [9]. Given a
feature point, the descriptor computes the gradient vector for each pixel in the feature point’s neighborhood and builds a
normalized histogram of gradient directions. The descriptor creates a neighborhood that is partitioned into sub-regions of
4×4 pixels each. For each pixel within a sub-region and adds the pixel’s gradient vector to a histogram of gradient directions
by quantizing each orientation to one of 8 directions and weighting the contribution of each vector by its magnitude.
Principle features of our scheme are summarized as Communication Efficiency: camera network is particularly well-suited
for low bandwidth; and unsupervised: the method does not require the pre-calibration into the scene and, hence, can be used
in traffic scenes where the system administrator may not have control over the activities taking place. Fig.5 shows the
matching results using descriptor created for a corresponding pair of points.
Fig 5 Matching results using descriptor.
We have experimented with various feature detectors including the Harris corner detector (HCD), curvilinear
structure detector (CSD), and difference of Gaussian (DoG) scale space. In Fig.6, the experimental result contain the
comparison of these methods is shown. We showed that suing SIFT point descriptors in a camera network can improves the
performance with respect to the other calibration systems. Here it is shown that descriptor lead to excellent performances
compared to other existing approaches. As explained, description is computed as follows: once a key-point is located and its
scale has been estimated, one or more orientations are assigned to it based on local image gradient direction around the keypoint. Then, image gradient magnitude and orientation are sampled around the key-point, using the scale of the key-point to
select the level of Gaussian blur. The gradient orientations obtained are rotated with respect to the key-point orientation
previously computed. Finally, the area around the key-point is divided in sub-regions, each of which is associated an
orientations histogram weighted with the magnitude.
Table1: Number of matching by features descriptors.
True Positive & False Positive
S1 Results
S2 Results
( number of occlusion: 31)
( number of occlusion: 23)
seconds Positive Positive seconds Positive Positive
In table2 counting and classification results are presented. As shown, the overall accuracy is about 90% for using
DOG detector in counting cars and about 94% for Bus and Trucks. This system can be as an input to calibration system in
multi-camera surveillance system.
Table2. Counting and classification results
Number of object matching by algorithm
Count Video Acc. Count Video
In this paper we considered the problem of features matching in a camera network with overlapping fields of view.
We showed that suing SIFT point descriptors in a camera network can improves the performance with respect to the other
calibration systems. In particular it returned good results for scale changes, zoom and image plane rotations, and large
1845 | Page
International Journal of Modern Engineering Research (IJMER)
Vol. 3, Issue. 3, May - June 2013 pp-1843-1846
ISSN: 2249-6645
point variations. These conclusions are supported by an extensive experimental evaluation, on different scenes. Therefore,
tracking and recognition using SIFT becomes feasible. This should result in highly robust trackers.
This work was supported by Islamic Azad University, West Tehran Branch.
Z.Fu, W.Hu and T.Tan, “Similarity Based Vehicle trajectory clustering and Anomaly Detection”, In Proc. Intl. Conf. on Image Processing (ICIP’5),
Vol 2, pp.602-605, 2005.
T.V.Mathew and K.V.Krishna Rao, “Introduction to Transportation Engineering.” Chapter 39,’Traffic Intersections’, NPTEL, 2007.
P.Babaei and M.Fathy,”Multi-Camera Systems Evaluation in Urban Traffic Surveillance versus Traditional Single-Camera Systems for Vehicles
Tracking,” 3rd International Conference on Computer and Electrical Engineering (ICCEE 2010), vol.4, pp.438-442, 2010.
P.Babaei and M.Fathy, “Vehicles tracking and classification using traffic zones in a hybrid scheme for intersection traffic management by smart
cameras”, In IEEE Conf. on Signal and Image Processing, (ICSIP 2010), pp. 49–53, 2010.
T.Bouwmans, F.E.Baf and B.Vachon, “Background modeling using mixture of Gaussians for foreground detection: a survey”, In proc of Patents on
Computer Science 1, pp.219–237, 2008.
M.A.Najjar, S.Ghosh and M.Bayoumi, “ A Hybrid adaptive Scheme based on selective gaussian modeling for real time object detection”, In IEEE
Symp. Circuit and systems, pp.936-939, 2009.
S.S.Cheung and C.Kamath, “Robust techniques for background subtraction in urban traffic video,” EURASIP J.Applied Signal Processing, vol. 5,
pp. 2330-2340, 2005.
E.B.Ermis, P.Clarot, P.Jodoin, “Activity Based Matching in Distributed Camera Networks,” IEEE Transaction on Image Processing, vol. 19, no.
10, OCT. pp.2595–2613, 2010.
D.Devarajan, Z.Cheng, and R Radke, “Calibrating distributed camera networks,” Proc. IEEE, vol. 96, no. 10, pp. 1625–1639, OCT. 2008.
C.Harris and M.Stephens, “A combined corner and edge detector,” in Proc. of 4th Alvey Vision Conf., pp. 147–151, 1988.
D.Lowe,“Distinctive image features from scale-invariant keypoints,”In Int. J. Comput. Vis., vol.60, no.2, pp. 91–110, 2004.
B.Song and A.R.Chowdhury,“Stochastic adaptive tracking in a camera network,” in Proc.IEEE Int.Conf.Computer Vision, pp.1–8, 2007.
S.Khan and M.Shah,“Tracking multiple occluding people by localizing on multiple scene planes,”IEEE Trans. Pattern Anal. Mach. Intell.,vol. 31,
no. 3 , pp. 505–519, Mar. 2009.
D.G.Lowe, “Object recognition from local scale-invariant features,” in Proc. of ICCV, pp. 1150–1157, 1999.
K.Mikolajczyk, C.Schmid,”A performance evaluation of local descriptors,” in Proc. Of CVPR, 2003, pp. 257–263.
K.Mikolajczyk and C.Schmid, “Indexing based on scale invariant interest points,” in Proc. Of ICCV, pp. 525-531, 2001.
T.Lindeberg, “Feature detection with automatic scale selection,” in Proc. Of IJCV, vol. 30, no.2, pp.79-116, 1998.
K.Mikolajczyk, A.Zisserman,“Shape recognition with edge-based features,” in Proc. of the British conf. Machine Vision 2003.
S.Belongie, J.Malik and J.Puzicha, “Shape context: A new descriptor for shape matching and object recognition,” in Proc. of NIPS, pp. 831-837,
A.Baumberg,”Reliable feature matching across widely separated views,” in Proc. of CVPR, pp.774– 781, 2000.
N.Allezard, M.Dhome, F.Jurie,”Recognition of 3D textured objects by mixing view-based and model based representations,” in Proc. of ICPR,
J.Matas, O.Chum, M.Urban, T.Pajdla,”Robust wide baseline stereo from maximally stable external regions,” in Proc. of BMVC, pp. 384–393,
K.Mikolajczyk, C.Schmid,”Scale and affine invariant interest point detectors,”In Int. Com. Vis. Vol. 60, no.1, pp. 63–86, 2004.
L.Lee, R.Romano, and G.Stein,“Monitoring activities from multiple video streams: Establishing a common coordinate frame,” IEEE Trans.Pattern
Anal.Mach Intell., vol.22, no.8, pp.758–767, Aug. 2000.
X.Wang, K.Tieu, and E.Grimson,“Correspondence-free activity analysis and scene modeling in multiple camera views,” IEEE Trans. Pattern Anal.
Mach. Intell., vol.32, no.1, pp.56–71, Jan. 2010.
D.Makris, T.Ellis, and J.Black, “Bridging the gap between cameras,” in Proc. of CVPR, vol. 2, pp. 205–210, 2004.
T.Chang and S.Gong, “Tracking multiple people with a multi-camera system,” in Proc. Of IEEE Multi-Object tracking, pp. 19–26, 2001.
1846 | Page
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF