A mobile ground-based radar sensor for detection and tracking of moving objects
© Vivet et al; licensee Springer. 2012
Received: 14 May 2011
Accepted: 23 February 2012
Published: 23 February 2012
The detection and tracking of moving objects (DATMO) in an outdoor environment from a mobile robot are difficult tasks because of the wide variety of dynamic objects. A reliable discrimination of mobile and static detections without any prior knowledge is often conditioned by a good position estimation obtained using Global Positionning System/Differential Global Positioning System (GPS/DGPS), proprioceptive sensors, inertial sensors or even the use of Simultaneous Localization and Mapping (SLAM) algorithms. In this article a solution of the DATMO problem is presented to perform this task using only a microwave radar sensor. Indeed, this sensor provides images of the environment from which Doppler information can be extracted and interpreted in order to obtain not only velocities of detected objects but also the robot's own velocity.
Keywordsradar sensor detection tracking moving objects
The detection and tracking of moving objects (DATMO) are among the most challenging problems concerning autonomous driving in a dynamic environment. Although the DATMO problem has been extensively studied for decades [1–6], it is still very difficult to accomplish these tasks from a ground vehicle at high speeds in outdoor environments.
Indeed, the most difficult issue is to separate moving objects from stationary objects. A classical approach in indoor environments is to use appearance-based or feature-based techniques with cameras and laser [7–9]. Both methods rely on prior knowledge of the targets. In an outdoor context, there are many types of mobile objects such as pedestrians, animals, vehicles of different sizes (cars, trucks, etc.), which are all very difficult to detect and identify.
Furthermore, in an outdoor environment DATMO is more complex under various climatic constraints. In this context, classical sensors are limited due to the technologies used: ultrasound is perturbed by wind, optical sensors (laser, vision) by rain, fog or the presence of dust or by poor lighting conditions. One of the particularities of this work is the use of a microwave radar sensor. In our case, the information from this sensor regarding the signal power reflected by the targets with a 360° per second rotating antenna and a range from 5 to 100 m is used. The long range and the robustness of radar waves to atmospheric conditions make this sensor well suited for extended outdoor SLAM and DATMO applications.
In classical detection and tracking approach, multiple detections of a same object have to be done in order to obtain target velocity. Each potential moving object is tracked and its model representing both position and speed is initialized and updated. In such a method, the large number of false tracks launched, combined with incorrect data association lead to algorithm failures. In this article, based on the radar frequency modulation principle, the sensor is able to provide two simultaneous images of the environment from which Doppler information can be extracted. Also, both the distance and velocity of the targets can be estimated simultaneously. It allows to create and initialize tracks at the first detection reducing following false data association.
In Section 2 a review of articles related to our research work is carried out. Section 3 briefly presents the microwave radar scanner developed by a Cemagref Institute team working on environmental sciences and technologies. Section 5 gives the principle used in this article in order to estimate the robot's own velocity. Sections 6 and 7 present, respectively the DATMO. Finally Section 8 shows experimental results of this work. Section 9 concludes.
2 Related work
In most applications, in order to accomplish DATMO from a mobile platform, an accurate localization systems are essential [10, 11]. Unfortunately, inertial measurement system is often very expensive and Global Positionning System (GPS) or Differential Global Positioning System (DGPS) often fail in an urban or covered environment such as forests because of the canyon effect. In the past decade, the simultaneous localization and mapping (SLAM) problem has been intensively studied in robotics because it can provide an accurate estimate of the robot position without expensive inertial sensors or GPS and it allows to build consistent map of the surroundings without prior knowledge. For a broad and quick review of the different approaches developed to address this problem, the reader can consult the following articles [12–15].
Most of the existing SLAM methods assume that the environment is static. If there is a moving object, and the data is erroneously associated with a landmark in the map database, many localization algorithms will fail, and the map will be deteriorated by the data of the moving object. The key point to solve this problem is to isolate the data of moving and static objects. Wang presented an approach to tackle SLAM and DATMO problems and proved that both problems are mutually beneficial . These two research areas are studied jointly under the denotation SLAM and Moving Object Tracking (SLAMMOT).
In order to deal with dynamic objects, Hahnel et al.  filtered out moving people, and created a difference map between consecutive laser scans to remove those static but people-like objects. An implicit assumption here is that dynamic objects move all the time during their measurements. However this is not normally true. Wang  did on-line calculations of an occupancy map, and detected the objects that entered an object-free space. More recently, Xie et al. , developed a SLAMMOT application based on probabilistic occupancy grids.
In order to perform outdoor SLAM or SLAMMOT, laser sensors are widely used [15–20]. Research work will continue to use them due to the success story of the Velodyne HDL-64 3D LIDAR . Visual sensors are also used to solve SLAMMOT problems. Ess et al.  presented an approach of multiperson tracking using a stereo rig mounted on a mobile platform. Solà et al.  described a system based on a framework called BiCamSLAM, that combines the advantages of monocular reconstruction with the advantages of stereo vision. Marzorati et al.  showed that the problem of SLAMMOT can be solved with a single camera.
In the naval field , the use of radar sensor for SLAM(MOT) application is self-evident but in ground mobile robotics few works use such a kind of sensor. In , we described a trajectory-oriented SLAM. It is based on radar information over important distances using Fourier-Mellin transform for scan matching considering a static environment. Radar is an interesting sensor because not only range and bearing can be obtained but also Doppler information can be used to extract velocities. This Doppler information allows to relax the assumption of static environment and to extend Radar SLAM to a SLAMMOT algorithm. In classical applications, successive acquisitions are compared knowing the localization (or computing it) in order to have an idea of the movements in the surroundings of the vehicle. In our proposition, as Doppler information is measured, we do not have to wait for two successive observations to obtain detection velocities. As a result, the estimation of ego-motion from different sensor acquisitions or proprioceptive sensors is not needed. In our case, the DATMO problem can be solved without SLAM information considering moving objects in the radar frame due to the fact that Doppler is measured directly. In this article a DATMO algorithm based on Doppler information is described using the IMPALA radar sensor.
3 The IMPALA radar
3.1 Range resolution
This expression of the range resolution is a theoretical relationship, it assumes a perfect linear modulation of the transmitted signal.
3.2 Velocity resolution
3.3 IMPALA radar characteristics
Characteristics of the IMPALA radar
Transmitter power Pt
Antenna gain G
Carrier frequency F 0
24.125 GHz (K band)
Angular resolution (horizontal)
Distance resolution δR
Velocity resolution δV
4 Issues of DATMO using a mobile ground-based radar sensor
In order to tackle the DATMO problem with our ground-based radar sensor, different problems have to be analyzed and solved. Before detecting moving objects and estimating their speed, the Doppler effect created by the vehicle's own velocity has to be estimated. In this step, the Doppler disturbance created by the vehicle itself has to be removed from the radar data. Next, in order to extract non coherent entities, both corrected images obtained from the up and down modulations are compared. Differences between the scans indicate potential moving objects. As the radar is subjected to important noises detected as differences between up and down images, false detections occur and have to be filtered out. Once moving objects are detected, a tracking process can be launched. Each moving object detection is compared and associated to the list of existing moving objects in order to update or create a new track. The approach that is used here is based on a classical Kalman process. The choice of Kalman filter does not affect the reliability of our solution, even though we are aware that better alternatives could be used, especially when dealing with the problem of data association [2, 29, 30]. But our goal in this article was to focus our study on the behavior of a DATMO algorithm based on the Doppler information in a ground-based radar environment, using a well-known filter to make correct conclusions.
In the remaining part of this article we detail the process of vehicule's own velocity estimation (cf. Section 5). In Section 6, extraction and filtering steps of non-coherent entities considered as mobile objects are explained. Finally Section 7 presents the tracking methodology and all the experimental results are discussed in Section 8.
5 Robot velocity estimation
where ○ is the Hadamard product function.
Each measurement of Doppler velocity VDoppler i has an uncertainty σDoppler. As a result, parameters of X of the function V(t) are estimated with their own uncertainty. Vehicle's own velocity profile V(t) and uncertainty σV(t)can be known during the radar acquisition.
6 Search of non coherent entities
As radar is subjected to important noises detected as differences between up and down images, false detections occur. A correlation score is used in order to filter some of false detections. Other false detections will be filtered out by the temporal moving object tracking process and the probabilistic approach (see Section 7).
In order to explain these true positive (TP) and false negative (FN) rates, radar and Doppler characteristics need to be considered. Doppler represents the radial velocity of objects. When an object is moving perpendicularly to the sensor, Doppler is null and so there is no detection. This explains the 20% FN rates at low range. Moreover, because of radar signal properties, detections at a high range are less powerful and much more noised than detections at a short distance. So the longer the range of the detection is, the lower the TP rates of moving object detection are.
At the end of this detection step, each potential moving object detected (noted O) is initialized as follows: O = [X o , V o , p o ] where X o = (x o , y o ) is the position of the object in the radar frame, is the object's velocities and p o is the probability of being a mobile object. This probability is obtained based on the detector characterization and varies according to the distance from the radar.
7 Tracking of moving objects
Each moving object detection is compared and associated to the list of existing moving objects in order to update or create a new track. This Detection association is based on the classical Mahalanobis distance taking into account both position and Doppler measurement along with their uncertainties. For each potential mobile object, tracking is done with a classical Kalman approach based on a constant velocity model. Other tracking methods using Interacting Multiple Model (IMM) and Multiple Hypothesis Tracking (MHT) techniques could be used to refine detection and data association . Additional difficulty with radar sensor is the absence of shape information.
with p(O) the prior probability of the track, p(d|O) and the TP and FN rates of the detector respectively. These rates are linked to the distance of the detected object (cf. Figure 8). Then posterior probability p(O|d) or becomes the new object existence probability p(O).
Track management is done based on different criteria: in case of out of range moving objects (> 100 m) or low probability of existence (p o < 0.05).
8 Experimental results
For these experiments, two experimental vehicles have been used. One was equipped with proprioceptive sensors, D-GPS for ground truth estimation and IMPALA FMCW panoramic radar imager. The other one called Vélac acts as the target (cf. Figure 3), and was equipped with proprioceptive sensors and D-GPS as well to have ground truth for moving object detection and tracking. Experiments were conducted in Clermont-Ferrand, France, on Blaise Pascal University campus, at variable speeds (with maximum 30 km/h).
8.1 Robot's own velocity estimation
Doppler velocity estimation with correlation presents a standard deviation of 0.3 m/s which corresponds to the correlation resolution. The estimated speed with its respective uncertainty is presented in Figure 10. A statistical evaluation of our Doppler odometry has been done. The linear velocity estimate error ϵ V has a standard deviation and a mean . An error during the classical odometer recording occurred at the end of the trajectory, which explains the 0 values on the red data while Doppler is still estimating the velocity.
8.2 Detection and tracking of moving objects
Trajectories presented in Figure 13 represent all the launched tracks. Two of them are due to real moving objects, while the remaining tracks are due to noises. Nevertheless, even if noise is important, their probability of existence is always decreasing and after five acquisitions (indeed 5 s) the majority of them are deleted as considered disturbances, while the remaining are confirmed as real mobile objects.
A method based on Doppler measurements for computing position and instantaneous velocity of moving objects in the surroundings of a robot using an original panoramic radar sensor was presented. Our IMPALA radar uses LFMCW in order to obtain both the radar-target distance and radial velocity of the target. With such a kind of ground-based radar sensor, the extraction and processing of landmarks remain a challenge because of detection ambiguity, false detection, Doppler Speckle effect and the absence of detection descriptors. Moreover, the data is affected by the Doppler effect created by the vehicle's own velocity. Correction based on a Doppler velocimetry has been applied in order to globally correct radar data. Once data is free from radar movement disturbances, non-coherent radar echoes are extracted and supposed as new moving objects. The probabilistic evaluation of our detector has been done and used to confirm or invalidate launched tracks at each new detection. Tracking of each entity is based on a classical extended Kalman filter. This approach was evaluated on real radar data, first, showing exteroceptive Doppler velocimetry feasibility and reliability at high speed (≈ 30 km/h), then detecting a D-GPS referenced moving object in a very noisy environment. A comparison between radar DATMO results and ground truth has been done. The main novelties of the proposed approach are the use of a panoramic LFMCW radar sensor and Doppler information for a ground mobile robotic application for DATMO purpose. Future work will include improving our radar SLAM (SLAM) process  by adding consideration of distortion due to non instantaneous data acquisition, Doppler information, and, as a consequence, a DATMO algorithm to tackle radar SLAMMOT problems in an extended outdoor environment. Moreover implementation of other filter techniques such as GM-CPHD (Gaussian mixture cardinalized probability hypothesis density)  will be compared with the actual Kalman method.
This study was supported by the Agence Nationale de la Recherche (ANR--the French national research agency) (ANR Impala PsiRob--ANR-06-ROBO-0012).
- Bar-Shalom Y: Tracking methods in a multitarget environment. IEEE Trans Automat Control 1978, 23: 4.Google Scholar
- Bar-Shalom Y, Li XR: Multitarget-Multisensor Tracking: Principles and Techniques. YBS, Dan-vers, MA; 1995.Google Scholar
- Blackman S, Popoli R: Design and Analysis of Modern Tracking Systems. Artech House, MA; 1999.Google Scholar
- Blom H, Bar-Shalom Y: The interacting multiple model algorithm for systems with Markovian switching coefficients. IEEE Trans Automat Control 1988, 33(8):780-783. 10.1109/9.1299View ArticleGoogle Scholar
- Reid DB: An algorithm for tracking multiple targets. IEEE Trans Automat Control 1979, 24(6):843-854. 10.1109/TAC.1979.1102177View ArticleGoogle Scholar
- Schulz D, Burgard W, Fox D, Cremers AB: Tracking multiple moving targets with a mobile robot using particle filters and statistical data association. In IEEE Int Conf on Robotics and Automation. Seoul, Korea; 2001:1665-70.Google Scholar
- Kluge B, Kohler C, Prassler E: Fast and robust tracking of multiple objects with a laser range finder. In IEEE Int Conf on Robotics and Automation. Seoul, Korea; 2001:1683-88.Google Scholar
- Lindstrom M, Eklundh JO: Detecting and tracking moving objects from a mobile platform using a laser range scanner. In Proc Int Conf On Intelligent Robots and Systems. Maui, HI, USA; 2001:1364-69.Google Scholar
- Gidel S, Checchin P, Blanc C, Chateau T, Trassoudaine L: Pedestrian detection and tracking in an urban environment using a multilayer laser scanner. IEEE Trans Intell Trans Syst 2010, 11(3):579-588.View ArticleGoogle Scholar
- Prassler E, Scholz J, Fiorini P: A Robotic Wheelchair for crowded public environments. Robot Automat Mag 2001, 8: 38-45. 10.1109/100.924358View ArticleGoogle Scholar
- Zhao L, Thorpe C: Qualitative and quantitative car tracking from a range image sequence. In IEEE Conf on Computer Vision and Pattern Recognition. Santa Barbara, CA, USA; 1998:496-501.Google Scholar
- Bailey T, Durrant-Whyte H: Simultaneous localization and mapping: part II--state of the art. Robot Aut Mag 2006, 13: 108-117.View ArticleGoogle Scholar
- Dissanayake G, Newman P, Durrant-Whyte H, Clark S, Csobra M: A solution to the simultaneous localization and map building problem. IEEE Trans Robot Autom 2001, 17(3):229-241. 10.1109/70.938381View ArticleGoogle Scholar
- Durrant-Whyte H, Bailey T: Simultaneous localization and mapping: part I--the essential algorithms. Robot Autom Mag 2006, 9: 99-108.View ArticleGoogle Scholar
- Wang CC: Simultaneous Localization, Mapping and Moving Object Tracking. In Ph.D. thesis. Carnegie Mellon Univ; 2004.Google Scholar
- Hahnel D, Burgard W, Fox D, Thrun S: An efficient FastSLAM algorithm for generating maps of large-scale cyclic environments from raw laser range measurements. In Proc Conf on Intelligent Robots and Systems. Las Vegas, USA; 2003:206-211. 2003,Google Scholar
- Xie J, Nashashibi F, Parent M, Garcia Favrot O: A real-time robust SLAM for large-scale outdoor environments. 17th ITS World Congress 2010.Google Scholar
- Pfaff P, Triebel R, Stachniss C, Lamon P, Burgard W, Siegwart R: Towards mapping of cities. In Proc of the IEEE Int Conf on Robotics and Automation (ICRA). Rome, Italy; 2007:4807-4813.Google Scholar
- Howard A, Wolf D, Sukhatme G: Towards 3D mapping in large urban environments. In Proc IEEE/RSJ Int Conf on Intelligent Robots and Systems (IROS). Sendai, Japan; 2004:419-424.Google Scholar
- Bosse M, Zlot R: Map matching and data association for large-scale two-dimensional laser scan-based SLAM. Int J Robot Res 2008, 27: 667-691. 10.1177/0278364908091366View ArticleGoogle Scholar
- Leonard J, How J, Teller S, Berger M, Campbell S, Fiore G, Fletcher L, Frazzoli E, Huang A, Karaman S, Koch O, Kuwata Y, Moore D, Olson E, Peters S, Teo J, Truax R, Walter M, Barrett D, Epstein A, Maheloni K, Moyer K, Jones T, Buckley R, Antone M, Galejs R, Krishnamurthy S, Williams J: A Perception Driven Autonomous Urban Vehicle. J Field Robot 2008, 25(10):727-774. 10.1002/rob.20262View ArticleGoogle Scholar
- Ess A, Leibe B, Schindler K, Gool LV: Robust multi-person tracking from a mobile platform. IEEE Trans Pattern Anal Mac Intell 2009, 31(10):1831-1846.View ArticleGoogle Scholar
- Solà J, Monin A, Devy M: Bicamslam: Two times mono is more a than stereo. In Proc of the IEEE Int Conf on Robotics and Automation (ICRA). Roma, Italy; 2007:4795-4800.Google Scholar
- Marzorati D, Matteucci M, Migliore D, Rigamonti R, Sorrenti G: Use a single camera for simultaneous localization and mapping with mobile object tracking in dynamic environments. In ICRA09 Workshop on Safe navigation in open and dynamic environments Application to autonomous vehicles. Kobe, Japan; 2009.Google Scholar
- Bibby C, Reid I: A hybrid SLAM representation for dynamic marine environments. In IEEE Int Conf on Robotics and Automation ICRA. Anchorage, Alaska, USA; 2010:257-264.Google Scholar
- Gérossier F, Checchin P, Blanc C, Chapuis R, Trassoudaine L: Trajectory-oriented EKF-SLAM using the Fourier-Mellin transform applied to microwave radar images. In IEEE/RSJ Int Conf on Intellig Robots and Systems (IROS). St. Louis, USA; 2009:4925-4930.Google Scholar
- Skolnik M: Introduction to Radar Systems. McGraw Hill, New York; 1980.Google Scholar
- Reiher M, Yang B: Derivation of the frequency mismatch probability in linear FMCW radar based on target distribution. In IEEE RadarCon 2009. Pasadena, USA; 2009:1-6.Google Scholar
- Ulmke M, Erdinc O, Willett P: GMTI tracking via the gaussian mixture cardinalized probability hypothesis density filter. IEEE Trans Aerosp Electron Syst 2010, 46: 1821-1833.View ArticleGoogle Scholar
- Mahler R: PHD filters of higher order in target number. IEEE Trans Aerosp Electron Syst 2007, 43: 1523-1543.View ArticleGoogle Scholar
- Press WH, Teukolsky SA, Vetterling WT, Flannery BP: Numerical Recipes 3rd Edition: The Art of Scientific Computing. 3rd edition. Cambridge University Press, New York; 2007.Google Scholar
- Ulmke M, Erdinc O, Willett P: GMTI tracking via the gaussian mixture cardinalized probability hypothesis density filter. IEEE Trans Aerosp Electron Syst 2010, 46: 1821-1833.View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.