Model-based Detection and Tracking of Single Moving Object using Laser Range Finder

2014 Fifth International Conference on Intelligent Systems, Modelling and Simulation Model-based Detection and Tracking of Single Moving Object using...
Author: Allan Neal
2 downloads 0 Views 473KB Size
2014 Fifth International Conference on Intelligent Systems, Modelling and Simulation

Model-based Detection and Tracking of Single Moving Object using Laser Range Finder Abdul Hadi Abd Rahman∗ , Hairi Zamzuri† , Saiful Amri Mazlan∗ and Mohd Azizi Abdul Rahman∗ ∗

Vehicle System Engineering Research Lab Universiti Teknologi Malaysia, Jalan Semarak, Kuala Lumpur Email: [email protected], [email protected], [email protected] † Corresponding Author UTM-Proton Active Safety Lab Universiti Teknologi Malaysia, Jalan Semarak, Kuala Lumpur Email: [email protected]

applied by [3], [4] and [5] are model-free approaches which have the ability to detect any type of dynamic object with no a prior knowledge about the object. Flexible models was proposed by [6] to detect moving vehicles. They introduced a method of constructing a virtual grid in a polar coordinate from laser data and use a scan differencing technique to detect motion evidences. Flexible rectangular models are matched to these evidences and vehicle sizes can be determined after several observations. This method can only detect vehicles successfully but does not apply for pedestrians, bicyclists or motorcyclists. An enhancement has been done by [7] which introduced another model-based approach which defined fixed models to represent several typical moving object classes. The method has an ability to perform moving object detection and tracking for buses, cars, motorcycle and pedestrians. This method showed the importance of using a geometric vehicle model which allows to naturally handle the disjoint point clusters and the estimation of geometric shape of vehicles leads to more accurate tracking results. A modelbased approach is utilized to eliminate the disadvantages of object segmentation when using laser scanners for tracking. Models were used to detect and classify dynamic objects for accurate tracking results. Several methods have been applied to solve moving object tracking. Aycard et al. [8] implemented a Markov Chain Monte Carlo technique to solve both detection and tracking by searching for the optimum solution in the spatio-temporal space. The implementation of bus model in solving separate segments of a detected bus contributed to successful object identification and tracking. A simulation on Fast-IMM (Fast-Interacting Multiple Model) has been done by [9] for tracking a single maneuvering target. The simulation results showed that the method decreased not only the computational burden but also keeps a high accuracy which is important for real time implementation. An Extended Kalman Filter (EKF) was used in [10] for

Abstract—Autonomous vehicles navigation in urban area suppose to deal with static and dynamic objects. To ensure safe autonomous navigation, the detection accuracy of a moving object is very critical to avoid false alarm while tracking the direction of moving object for collision avoidance. This paper presents the experimental results of the proposed Detection and Tracking Moving Object (DATMO) algorithm for a single moving object. Experiments were conducted in outdoor environment using a vehicle equipped with a laser range finder to provide the input data to the DATMO algorithm while the Kalman filter with the integration of constant velocity (CV) model was used for tracking. The results from the experiments showed that the DATMO algorithm developed successfully tracks a single moving object in an outdoor environment. Keywords-DATMO; target tracking; autonomous vehicle; intelligent system;

I. I NTRODUCTION One of the most important capabilities of an Autonomous Guided Vehicle (AGV) is to detect and track moving objects. With this feature, AGV will be able to avoid obstacles such as cars, pedestrian and other dangerous situations while travelling along the pathway. The moving object tracking problem has been originated from radar tracking systems and extensively studied for several decades [1]. Detection and Tracking Moving Objects (DATMO) deal with dynamic objects detected from the environment. DATMO enables the prediction of the future behaviors of dynamic objects that can be useful for autonomous vehicle for safety reasons such as collision avoidance. Various methods have been developed to detect moving objects with a particular use of laser sensors. A method based on simple features has been developed to detect people in office environments. Indoor people can be distinguished by detecting local minimal in the laser range scans. This method is only practicable on the restricted ambient, since in an outdoor condition, a tree will be recognized as a pedestrian. Another weakness of this method is the difficulty to detect other object classes rather than people [2]. Methods 2166-0662/14 $31.00 © 2014 IEEE DOI 10.1109/ISMS.2014.102

556

each detected pedestrian to solve the occlusion problem and provides computational advantage for tracking in real time. Experiments were conducted to track pedestrian moving in various directions from a mobile robot. Further improvement on object tracking has been done by [11] using an adaptive Interacting Multiple Models filter which consists of 16 motion models coupled with a Multiple Hypothesis Tracker to solve moving objects tracking. The experimental results showed that the method successfully performed a real time moving object tracking from a vehicle at high speeds in different dynamic outdoor scenarios. This paper presents the experimental results of DATMO algorithm using model-based for object detection; and Kalman filter with constant velocity (CV) to represent the the motion model of the object movement for tracking a moving object. II. D ETECTION AND T RACKING M OVING O BJECT

Figure 1.

In computer vision, moving object detection algorithms can be classified as appearance-based, feature-based, motion-based and model-based methods [12]. In comparison with images, laser data contained less information thus the appearance-based approaches are not directly applicable. When moving objects are detected and located, it is desirable to track them in order to estimate their dynamic states which allow aggregation of object observations for estimation enhancement. The state vector includes position, speed, acceleration, geometric description of detected objects. Usually these state variables cannot be observed or measured directly, but they can be estimated through tracking process. The general recursive probabilistic formula for moving object tracking can be expressed as p(ok , sk |Zk )

III. M ETHODOLOGY A. Overview The flow of DATMO algorithm is illustrated in Figure 1. It begins with scan lines that can be directly obtained from a 2D LIDAR scanner. This scan line is transferred into world coordinates and segmented. Line and corner features are extracted for each segment. The segments are associated with existing objects and the kinematics of the objects is updated with the use of Kalman filter. Segments that cannot be associated could trigger the creation of a new object. The Kalman filter assumes the errors to be Gaussian applicable to the measurement sensor used in the study. B. Pre-Processing

(1)

The readings are subdivided into sets of neighbour points (clusters) by taking the proximity between each two consecutive points of the scan. A cluster is a set of measurements from the points of the scan which close enough to each other due to their proximity would probably belong to the same object. The clustering algorithm is given as follows:

where ok is the true state of a moving object at time k, and sk is the true motion mode of the moving object at time k, and Zk is the perception measurement set leading up to time k. The robot (sensor platform) is assumed to be stationary for the sake of simplicity. Using Bayes rule, Equation (1) can be rewritten as p(ok , sk |Zk ) = p(ok |sk , Zk ) . p(sk |Zk )

Flow of DATMO algorithm.

rAB =

(2)

 2 2 rOA + rOB − 2rOA rOB cosα

rAB ≤ C0 + C1 min{rOA , rOB }  c1 = 2(1 − cosα)

which indicates that the moving object tracking problem can be solved in two stages: the first stage is the mode learning stage p(sk |Zk ) and the second stage is the state inference stage p(ok |sk , Zk ). In practice, the motion mode of moving objects can be approximately composed of several motion models such as the constant velocity model, the constant acceleration model and the turning model. However, in this paper only constant velocity is considered as the motion model for tracking the moving object.

(3) (4) (5)

Assuming that the moving object can be approximated by polygonal shapes, line fitting is a suitable choice for object faces approximation. Later, all the broken objects are joining during the grouping and line fitting processes as illustrated in Figure 2. The algorithm searches for the point Pj with the greatest distance from the line through P0 and Pn . If the distance is greater than the threshold T , the line P0 Pn

557

Figure 2.

Iterative End Point (IEPF) algorithm for line fitting process.

is broken into two lines P0 Pj and Pj Pn , and the process repeats for the two lines. Objects should own different features such as type, velocity, size. Objects can be classified into predefined types which have been tracked from previous scans. Ideally, the object classification should give every object an accurate classification on every scan. However, due to the laser sensor characteristics, it is difficult to differentiate the shape of the same object. Thus, a pre-defined box model size for specific targeted object was used to determine the centroid of moving object for tracking.

Figure 3.

are, ν = (Zk − Zk|k−1 ) k|k−1 Zk|k−1 = Hk x

(12)

Sk = Hk Pk|k−1 HkT + Rk

(14)

Wk =

Pk|k−1 HkT Sk−1

(13) (15)

D. Motion Model

C. Kalman Filter

A constant velocity (CV) model was applied to capture the target trajectories where the target moves at almost constant velocities. The CV model is given as follows:      1 t x1 (k − 1) x1 (k = + 0 1 x2 (k) x2 (k − 1) ⎡ ⎤ t2 ⎣ 2 ⎦ v(k − 1) (16) t

The Kalman filter is an estimator for discrete-time linear dynamic systems driven by white noise. It was introduced in 1960s as the optimal recursive estimation in linear Gaussian models. In the Kalman filter framework, the motion and the measurement models are both linear and corrupted by zero mean Gaussian noises wk and vk , xk = Fk xk−1 + Gk uk − 1 + wk Zk = Hk xk + vk

(6) (7)

y(k) = [1 0]x(k) + ω(k)

where matrices Fk , Gk and Hk are time dependent and the noises wk and vk have covariance matrices Qk and Rk respectively. If the prior distribution p(xo ) is also a o , Po ), the distributions resulting from the Gaussian N (xo ; x prediction and update steps are also Gaussian. Therefore, the beliefs of the Kalman filter are completely specified by the first two moments of the distribution. Thus, the distribuk|k1 , Pk|k−1 ) where the tion p(xk |Z1:k1 , u0:k−1 ) = N (xk ; x mean xk|k−1 and covariance matrix Pk|k−1 are, k−1 + Guk−1 x k|k−1 = F x Pk|k−1 =

iREV Autonomous Vehicle.

Fk Pk|k−1 FkT

+ Qk

IV. E XPERIMENTAL S ETUP The experiments were conducted using a modified buggy golf cart which has been equipped with sensors and two onboard computers as shown in Figure 3. The trajectory data is obtained from a custom program developed by [13] from the Vehicle System Engineering (VSE) research group in Visual basic. In order to receive the real time data from the system, a communication channel using TCP/IP has been developed in the main algorithm. It has been validated and worked perfectly. A new algorithm for DATMO in C# has been developed to integrate the trajectory data and laser data from laser range finder. A laser range finder Hokuyo URG-04LX is placed in front of the vehicle (0.7m from ground) to provide 180o range of data within 4 meter radius as in Figure 4. The experiments have been done on a plane surface in outdoor environment. The experiments were conducted to evaluate the performance of the DATMO algorithm in 2 different situations which are:

(8) (9)

In the same way, the mean x k and covariance matrix Pk resulting from the update step are, k|k−1 + Wk νk x k = x Pk = Pk|k−1 −

Wk Sk WkT

(17)

(10) (11)

where the innovation ν, the predicted measurement Zk|k−1 , the innovation covariance Sk and the filter gain Wk

558

Figure 4.

(a) Graphic User Interface (GUI) for real-time data processing of laser range finder raw data.

Placement of laser range finder on the vehicle.

(b) Line extraction data. Figure 6.

Results of pre-processing of raw laser data

Figure 5.

Distance(cm)

40

Experiments for tracking single moving object.

Measured Box Model 50x15 Point Extraction

30 20 10

a) b)

Tracking a single moving object from a static vehicle. Tracking a dynamic single object which shall move, stop and resume to move again from a static vehicle.

0 −50

−40

−30

−20

−10

0

10

20

30

Distance(cm)

Figure 7.

V. R ESULTS AND D ISCUSSION

Point extraction using box model.

B. Box Model

This section presents the results of the data preprocessing, the box model, platform development and single moving object tracking. The experiments were conducted to track a single moving pedestrian within sensing area in front of the vehicle as in Figure 5.

Box model has been used as the method to extract the centroid point of the detected object. The point extraction for the box model illustrates in Figure 7. The box model size was predefined as 50cm by 15cm to represent the object size. The result of the point extraction was used in the following step for tracking purpose.

A. Clustering

C. Tracking

The clustered data from raw sensor readings are illustrated in real-time graph as in Figure 6(a). The graph shows that the algorithm successfully performed the data segmentation. Each cluster of data represents an object with various shapes. Later, a feature extraction process was performed to transform each cluster into a set of line according to the object feature. Figure 6(b) shows the details of extracted lines per cluster after the process where the first and second clusters contain a line whilst two lines are detected in the third cluster. The results suggested that the algorithm was successfully extracting the lines from the clustered data.

Kalman filter with CV model was used for tracking a moving object. Figure 8(a) shows the result of the tracking of moving pedestrian using Kalman algorithm from static vehicle located at (0,0) in the global coordinate. In this experiment, the pedestrian moved continuously in random direction which was then detected and tracked by the DATMO algorithm. The DATMO algorithm was able to predict the movement of the object correctly with the integration of constant velocity (CV) model. At the same time the algorithm produced the estimated movement of

559

160

140

140

120

120

100

100

Distance(cm)

Distance(cm)

160

80 60 40

80 60 40

20 0 −200

−100

0

100

200

Measured Kalman CV Vehicle Pose

20

Measured Kalman CV Vehicle Pose

300

−200

0

400

−150

−100

−50

Distance(cm)

100

100

50

50

0

−50

4

6

8

10

100

150

200

12

14

16

0

−50

Kalman dX Kalman dY 2

50

(a) Measured object position and Kalman with CV

Distance(cm)

Distance(cm)

(a) Measured object position and Kalman with CV

0

0 Distance(cm)

18

0

Iteration

Kalman dX Kalman dY 2

4

6

8

10

12

14

16

18

Iteration

(b) Vehicle position changes over time

(b) Vehicle position changes over time

Figure 8. Results of tracking a single moving object from a static vehicle.

Figure 9. Results of tracking a dynamic single object which move, stop and resume to move again from a static vehicle.

pedestrian over dX and dY as illustrated in Figure 8(b). Figure 8(b) shows that the detected and tracked object was travelling from left to right denoted by the negative value of dX and rapidly increasing towards positive values. The recorded values of dY which states almost constant at a value of 20cm shows the diagonal movement of the object in front of the vehicle. Figure 9(a) shows the result of the tracking of moving pedestrian using Kalman algorithm from static vehicle with different motion of the pedestrian. In this experiment, the pedestrian moved for a certain period of time, then stop and resume to move again after a while. At the same time the algorithm is able to produce estimated movement of pedestrian over dX and dY as illustrated in Figure 9(b). It shows that the object stopped for a certain period of time from 7th iteration to 11th and resumed to move to the right of the vehicle.

especially to solve the problem with various movement of tracking object especially on the multiple motion models. This algorithm will be extended for tracking single moving object from a moving vehicle and deals with multi moving objects. ACKNOWLEDGEMENT The authors would like to thank Malaysia-Japan International Institute of Technology (MJIIT) in UTM Kuala Lumpur, Ministry of Education Malaysia (MOE) and ERGS (ERGS code:4L033) for funding and supporting this research. R EFERENCES [1] H. Blom and Y. Bar-Shalom, “The interacting multiple model algorithm for systems with Markovian switching coefficients,” IEEE Transactions on Automatic Control, vol. 33, no. 8, pp. 780–783, 1988.

VI. C ONCLUSION AND F UTURE W ORK This paper mainly deals with tracking a single moving object using a single motion model. The tracking of moving object cannot be solved using Kalman filter alone but it needs to be integrated with a motion model to predict the dynamic movement of the object. The results show that the integration of Kalman filter with CV was able to solve tracking problem. A further improvement is needed

[2] D. Schulz, W. Burgard, D. Fox, and A. Cremers, “Tracking multiple moving targets with a mobile robot using particle filters and statistical data association,” in Proceedings 2001 ICRA. IEEE International Conference on Robotics and Automation (Cat. No.01CH37164), vol. 2. IEEE, 2001, pp. 1665–1670.

560

[9] D. Mohammed, K. Mokhtar, O. Abdelaziz, and M. Abdelkrim, “A new IMM algorithm using fixed coefficients filters (fastIMM),” AEU - International Journal of Electronics and Communications, vol. 64, no. 12, pp. 1123–1127, Dec. 2010.

[3] D. H¨ahnel, D. Schulz, and W. Burgard, “Mobile robot mapping in populated environments,” Advanced Robotics, vol. 17, no. 7, pp. 579–597, Jan. 2003. [4] C.-c. Wang, “Simultaneous Localization, Mapping And Moving Object Tracking,” Ph.D. dissertation, Carnegie Mellon University, 2004.

[10] C.-l. Chen, C.-c. Chou, and F.-l. Lian, “Active Pedestrian Following Using Laser Range Finder,” in IEEE Internatioinal Conference on Information and Automation, no. June, Shenzhen, China, 2011, pp. 690–695.

[5] L. Montesano, J. Minguez, and L. Montano, “Modeling the Static and the Dynamic Parts of the Environment to Improve Sensor-based Navigation,” Proceedings of IEEE International Conference on Robotics and Automation, pp. 4556–4562, 2005.

[11] T.-D. Vu, J. Burlet, and O. Aycard, “Grid-based localization and local mapping with moving object detection and tracking,” Information Fusion, vol. 12, no. 1, pp. 58–69, Jan. 2011.

[6] A. Petrovskaya and S. Thrun, “Model based vehicle detection and tracking for autonomous urban driving,” Autonomous Robots, vol. 26, no. 2-3, pp. 123–139, Apr. 2009.

[12] M. S. A. Ramli, H. Zamzuri, and M. S. Z. Abidin, “Tracking human movement in office environment using video processing,” in 2011 Fourth International Conference on Modeling, Simulation and Applied Optimization. IEEE, Apr. 2011, pp. 1–6.

[7] T.-D. Vu, “Vehicle Perception : Localization , Mapping with Detection , Classification and Tracking of Moving Objects,” Ph.D. dissertation, Grenoble Institute of Technology, 2009.

[13] M. Aizzat, H. Zamzuri, R. Mamat, and S. Amri, “A Path Tracking Algorithm Using Future Prediction Control with Spike Detection for an Autonomous Vehicle Robot,” International Journal of Advanced Robotic Systems, p. 1, 2013.

[8] O. Aycard, “Laser-based detection and tracking moving objects using data-driven Markov chain Monte Carlo,” 2009 IEEE International Conference on Robotics and Automation, pp. 3800–3806, May 2009.

561

Suggest Documents