Robust 3D Tracking of Multiple Objects with Similar Appearances(2)
1-Introduction
In 3D moving objects tracking is the trajectory of a moving object or 3D object in a sequence of input images. The three-dimensional moving objects that can be detected can be any three-dimensional moving object such as fish in the water, boats in the sea, pedestrians on the sidewalk, cars on the highways, etc., which we need to determine their location based on the application. For example, in an automated navigation system (driverless car) it should be possible to detect three-dimensional moving objects around the system and at any time determine the position of these objects to avoid colliding with other objects. Tracking algorithms must be able to withstand many challenges such as changes in ambient light intensity, sudden change in the direction of 3D objects, the presence of different types of 3D objects in the field of view of the camera, overlap, and so on.
2- Background
The 3D tracking literature is evolving day by day due to its wide and significant applications in the field of human security and comfort, so that it has emerged from the shadow of military applications and has shown its application in modern systems used by humans. In the articles, they have detected and tracked moving three-dimensional objects by updating the background model with a middle filter algorithm. At each step, they update the background by placing the middle value of the neighbors of each pixel, and by applying the subtraction algorithm to obtain the background of the moving areas, and then with the algorithm, discover the connected components of the moving three-dimensional objects and move from the center of each object to They used the title of tracking invoice. The proposed algorithms use many properties for tracking three-dimensional objects, such as edge properties, corner points to texture, color, and so on. Among the mentioned properties, corner properties have been widely used in tracking three-dimensional objects because corner points are more flexible than changes in ambient light intensity, change in the direction of moving objects, and change in angle of the tracked object.
3- Proposed Algorithm
In this algorithm, first using Sift corner points and matching these points with KLT algorithm and based on the movement information of key points related to moving objects, discover them and at each stage these clusters or three-dimensional moving objects on the received frames using similarity. The area of the object is obtained by averaging the brightness of the points belonging to that area.
3-1- Discover feature points related to three-dimensional moving objects with similar appearance
After comparing the SIFT feature points between two consecutive frames, the motion size (displacement rate) of the feature points can be calculated based on the Euclidean distance of the coordinates of the points for two consecutive frames in terms of time. To do this, formula (1) is used. Now from the corresponding feature points that correspond to two consecutive frames with the KLT algorithm. Label points that have acceptable displacement (key points) as three-dimensional moving objects in the tracking scene, and remove other points from the feature set. To do this, formula (2) is used.
3-1- Tracing Three-Dimensional Moving Objects
In the tracking step, we
have to determine the new locations of the 3D moving objects in the scene in
each received frame. In each iteration of the proposed algorithm and by
discovering three-dimensional moving objects in the camera field of view,
compare the value obtained for the average brightness of each object with the
values in the previous step and Based on a limit, we establish a one-to-one
correspondence threshold between three-dimensional moving objects. To do this,
formula (3) has been used. If the three-dimensional moving objects discovered
between the two frames are similar in terms of the average light intensity of
the area obtained for that object, then we consider a place as the new location
of the object that has the lowest value in terms of Euclidean distance. Have
several candidates among these. The results for tracking are shown in Figure
(2).
In Formula 3, N1 and N2 are the number of clusters in the previous
frame and the current frame, respectively, and Vi and Vj are the average light
intensities for moving objects detected at moments t-1 and t, respectively. In
this equation, X is the threshold for calculating the average light intensity.
If the average brightness of the objects detected in the field of view of the
camera in each repetition is higher than the proposed threshold, this object is
considered as a new three-dimensional object. Therefore, if a new 3D object
enters the tracking scene, the properties of this object will be saved and will
be tracked in subsequent iterations. Also, when an object leaves the scene
covered by the camera, since it is not among the newly discovered objects in
the next iteration, the tracking action on it is not considered automatically
until it re-enters the shooting scene. With this method, if moving objects with
different light intensities enter the tracking scene for reasons such as being
in the shadows, they will be tracked with high accuracy because in each
repetition, the representative of each object is updated (by averaging the
light intensity).
conclusions
Proposed method Sift features provides a
powerful matching despite changes in scale, rotation, noise, changes in
brightness, noise, and obstruction. The proposed algorithm uses the DBSCAN
density-based clustering algorithm in the moving object detection step, which
is performed on specific points related to similar three-dimensional moving
objects. The results of the proposed method are shown on three different videos
in Figure (3), which shows the proposed algorithm for tracking
three-dimensional moving objects similar to the type of objects, changes in
speed and direction of movement, and the existence of several three-dimensional
moving objects simultaneously in the field of view of the camera is not
sensitive.
References
[1] Minghua Li, Zhengxi Liu, Yunyu Xiong, and
Zheng Li. Multi-person tracking by discriminative affinity model and
hierarchical association. In 2017 3rd IEEE International Conference on Computer
and Communications (ICCC), pages 1741–1745. IEEE, 2017.
[2] Wenbo Li, Ming-Ching Chang, and Siwei Lyu.
Who did what at where and when: Simultaneous multi-person tracking and activity
recognition. arXiv preprint arXiv:1807.01253, 2018.
[3] Felipe Jorquera, Sergio Hernández, and
Diego Vergara. Probability hypothesis density filter using determinantal point
processes for multi object tracking. Computer Vision and Image Understanding,
2019.
[4] Zhao Zhong, Zichen Yang, Weitao Feng, Wei
Wu, Yangyang Hu, and Cheng-lin Liu. Decision controller for object tracking
with deep reinforcement learning. IEEE Access, 2019.
[5] J. Ning, L. Zhang, D. Zhang, and C. Wu, Robust
object tracking using joint color-texture histogram, International J. of
Pattern Recognition and Artificial Intelligence, vol. 23, no. 07, pp. 1245-
1263, Nov. 2019.
[6] Dawei Zhao, Hao Fu, Liang Xiao, Tao Wu,
and Bin Dai. Multi-object tracking with correlation filter for autonomous
vehicle. Sensors, 18(7):2004, 2018.
[7] Weigang Lu, Zhiping Zhou, Lijuan Zhang,
and Guoqiang Zheng. Multi-target tracking by non-linear motion patterns based
on hierarchical network flows. Multimedia Systems, pages 1–12, 2019.
[8] Nan
Ran, Longteng Kong, Yunhong Wang, and Qingjie Liu. A robust multi-athlete
tracking algorithm by exploiting discriminant features and long-term
dependencies. In International Conference on Multimedia Modeling, pages 411–423.
Springer, 2019.