Detection and Tracking Approach Using an Automotive Radar to Increase Active Pedestrian Safety

Abstract—Vulnerable road users, e.g. pedestrians, have a high impact on fatal accident numbers. To reduce these statistics, car manufactures are intensively developing suitable safety systems. Hereby, fast and reliable environment recognition is a major challenge. In this paper we describe a tracking approach that is only based on a 24 GHz radar sensor. While common radar signal processing loses much information, we make use of a trackbefore-detect filter to incorporate raw measurements. It is explained how the Range-Doppler spectrum of pedestrian can help to initiate and stabilize tracking even in occultation scenarios compared to sensors in series.


1. Introduction

In the policy orientations on road safety 2011-2020 the European Union (EU) sets itself the goal of halving the number of road accident victims by 2020 [1]. A decade before the same claim has not been met as advanced driver assistance systems (ADAS) were expected to catch on much faster than they have. But the high costs of the available ADAS are one reason why they are mainly in use in top-ofthe-range models. To have a significant effect on the reduction of fatality numbers, ADAS have to be economically viable for penetrating into the volume vehicle segment.

We identified three superior pedestrian accident types based on a German in-depth data base called GIDAS. From the available information, straight forward scenarios can be rendered, which occur most often during day-time in regular (dry, no dust) road conditions. Accident reasons are often the misbehavior of the driver and/or the pedestrian and an unadapt velocity, usually caused by distraction or misjudgement. One example for pedestrian misbehavior is a non-observance of traffic while stepping out of occlusion [2].

Among other initiatives to improve safety of vulnerable road users (VRUs), the European Commission (EC) is funding a research project called ARTRAC: “Advanced Radar Tracking and Classification for Enhanced Road Safety” aimed to develop an integrated safety concept for pedestrians. ARTRAC investigates the possibility to use a standalone radar sensor especially for the detection of VRUs in order to generate the appropriate supporting actions to the driver with the goal to avoid or mitigate a possible collision.

ARTRAC demonstrator vehicle equipped with high resolution radar sensor for pedestrian detection Figure 1: ARTRAC demonstrator vehicle equipped with high resolution radar sensor for pedestrian detection

Although, a few pedestrian safety systems are already available on the market [3, 4], consumer organization test have shown that their performance is limited even in a sensor-friendly situation. On the other hand all of these systems rely at least on a setup that merges two different sensor technologies. A standalone sensor solution using radar for the detection and tracking has currently several drawbacks to overcome. Compared to a multi sensor system the outcome still needs improvement to feasible estimate an adequate collision risk. For instance, a pedestrian has naturally a much lower radar cross section (RCS) as a passenger car, and thus is harder to identify in the received signal spectrum. Especially, in the proximity of cars, a correct distinction becomes intractable due to resolution capabilities and other radar related phenomena. This has also been stated by [5, 6], both describing a radar system for the purpose of object classification based on raw data. Tracking is considered as a high level task, without mentioning the achieved or expected precision.

While newly developed sensors increase data accuracy and resolution, conventional object detection methods are incapable of getting a maximum of information out of the data. In this paper we show how to use a track-before-detect approach (TBD), to consider measurements which are usually being discarded [7]. To handle the amount of data we chose a Particle Filter (PF) and the scenario-driven search approach SDS [8]. PF have been used extensively in a wide range of human tracking fields with satisfying results. Further on, a PF allows us to define a weighting that accounts Doppler signatures caused by human walking [9].

2. Radar Sensor

The radar sensor uses a carrier frequency within the ISM band from 24.00 GHz to 24.25 GHz. It has a multi receive antenna system and offers therefore a high resolution in azimuth direction. The transmitted waveform is a so called Rapid Chirp. It allows a simultaneous measure of range,relative radial velocity and azimuth angle of multiple targets. The waveform consists of L fast (up or down) ramps to process a 2D FFT that provides a Range-Doppler spectrum. During the duration of a single chirp the beat or difference frequency is dominated by the target range. Hence, the Doppler Effect can be neglected. But it causes a constant phase shift over consecutive ramps in relation to the radial velocity. After the target information is separated by range and speed the 2D FFT is mapped to several directions by means of DBF.

Further specifications are: a modulation bandwidth of 150 MHz, a sample rate of ~1.5 MHz and 256 fast ramps by 60 ms cycle time (alternating transmit beam +/-15°). The radar covers and area over up to 280 m and 60° with a resolution of 1 m and 6 [10].

A. Conventional processing of radar measurements
The general process to extract objects from radar raw data is depicted in the diagram of figure 2. Firstly, to reduce noise and find peaks, a decision threshold is calculated. Depending on the received signal strength, reflections above a certain level are further considered, otherwise the measurements are discarded. Methods to determine a suitable threshold are well known e.g. CFAR [11]. You can argue if those in some extent complex methods are able to detect targets with weak reflectivity like pedestrians. Afterwards clustering is performed to reduce the number of detections and eliminate multiple detections of single objects. In this process weak detections can mistakenly be assigned to intense reflections due to insufficient sidelobe compression. Thus, a significant loss of information is possible in both of these steps.

Finally, object states are extracted from the clusters and forwarded to a tracking algorithm. All valid tracks are stored in an object list that is available for subsequent applications. The details are explained in common radar compendia.


3. Micro doppler effect

From radar sensors’ point of view walking pedestrians are non-rigid bodies. Torso, arms and legs show an oscillating movement. In particular, the arm and leg movement causes Doppler measurements which differ from measurements of the torso.

The pedestrian is identifiable as an extended target in the radar raw data. As a result the Doppler signature of a walking pedestrian is clearly distinguishable from e.g. a vehicle or a static object. This signature is well described and evaluated in [9]. Figure 3 shows an example of measured Range-Doppler values for a single beam in which a pedestrian and a vehicle are departing from the radar sensor.
In contrast to the narrow Doppler-signature of the vehicle the micro Doppler Effect of the walking pedestrian causes an obvious widening of the signature. Although, this effect is primarily caused by radial movement, an attenuated signature is also visible when the pedestrian moves lateral through the sensors field of view.

Figure 2: Diagram of conventional signal processing tasks
Figure 2: Diagram of conventional signal processing tasks

Example of a Range-Doppler spectrum indicating a walking pedestrian and a passenger car

Figure 3: Example of a Range-Doppler spectrum indicating a walking pedestrian and a passenger car

4. Pedestrian tracking

A. Track-before-detect
To avoid the described loss of information by e.g. thresholding and clustering, a TBD approach has been chosen. Approaches in [12] and [13] show the ability for tracking weak signals in noisy measurements or besides strongly reflecting objects. Instead of using extracted objects the complete Range-Doppler domain for every digital beam can be used. Even so, it is possible to make use of the micro Doppler Effect and other features listed in [5] and [6], no expensive feature extraction is performed.

B. Particle Filter Preferences
Particle filters are known under different notations and are used in various areas. The main idea is to represent the posterior probability with a set of weighted particles . With this approach the posterior probability is formed by

Detection and tracking approach using an automotive radar to increase active pedestrian safety
with being the state hypothesis of particle i at timestep t.
Due to performance the set of particles sticks to a region of interest (ROI) similar as in [8]. The ROI is derived from vehicle dynamics and strong target reflections from static objects likely to come from cars. Initially, the particle states are randomly spread in this ROI with velocity values out of two intervals ranging from [-3, -1] and [1, 3] m/s, which is most likely for a walking pedestrians [5]. Prediction Model, Measurement Model and Resampling
After testing several predictions models, we figured out that the profit of a complex pedestrian movement model is negligible. Thus, we assume a constant velocity model for the short timeframe of a maximum of 300 ms (two consecutive measurements).

According to the measurement model the weights of each predicted particle state are updated by transforming the particle to a specific bin in the measured spectrum and calculate it’s Range-Doppler signature α.

Detection and tracking approach using an automotive radar to increase active pedestrian safety
This step causes a high weight for particles which are located in the area of a walking pedestrian and have a suitable Range-Doppler-profile. The next step is the resampling of the particle set. In this paper two thresholds are used to determine whether a particle should be duplicated, deleted or retained based on its weights. This method causes a clustering of particles when a pedestrian is walking through the region of interest shown in figure 4. This cluster concretes the object hypothesis represented by each particle and indicates a high probability for a walking pedestrian.

At the end of every filter cycle the object state for every tracked pedestrian has to be estimated from the given cluster. At this stage a preliminary cluster method has not been able to extract a plausible object state in every cycle.


5. Results

Measurements have been taken for different scenarios oriented on test procedures of consumer protection organization like Euro NCAP. The experimental setup to evaluate our approach consists of three sensors mounted on a demonstrator vehicle. Additionally to the described 24 GHz radar sensor (cycle time ~120 ms), the data of a 77 GHz radar (60 ms) and a mono-camera (250 ms) in series production are compared to a reference signal. A mobile DGPS system was carried by the pedestrian during all test runs which provides the reference signal while being wireless connected to the inertial measurement system of the demonstrator vehicle, a similar test system as described indepth by [14].

A safety relevant test case is illustrated in fig. 5, a pedestrian crossing the street in lateral direction. When start walking, the pedestrian is in the line of sight, then disappears behind a parking vehicle before stepping onto the driving lane. This situation is putting high requirements on the detection and tracking system in terms of reliability and robustness.

Schematic region of interest and particle cluster

Figure 4: Schematic region of interest and particle cluster


Parameter Sensors
24 GHz
77 GHz
Series Radar
- mean deviation
- standard deviation
- absolute max. deviation

0,13 m
0,91 m
1,79 m

-0,02 m
0,23 m
0,83 m

0,72 m
0,64 m
1,24 m
- mean deviation
- standard deviation
- absolute max. deviation

0,23 m/s
0,25 m/s
0,54 m/s

-0,06 m/s
0,18 m/s
0,57 m/s

0,84 m/s
2,69 m/s
9,78 m/s


Representatively, the lateral position as well as the longitudinal velocity of a walking pedestrian for all sensors is shown in in fig. 6 and 7. Notice that the mono-camera does not provide a lateral velocity.

The track results, made by the developed particle filter, have a higher standard deviation than the compared sensors, while we have used a small amount of 2000 particles. Remarkable is the sustained tracking during the occultation implied by the dots in the dashed area. Here, the TBD approach enables the usage of weak reflections through multipath propagation underneath the parked vehicle. Although, both radars clearly show measurements, the 77 GHz radar has not been able to track the pedestrian behind the car, neither does the mono camera. Due to the long time span a coasting mode can be excluded. It stands out that association is a more concerning issue. The measuring accuracy of the mono-camera alone has mostly been inacceptable for a pedestrian safety system. On the other hand the 77 GHz is overall providing the best results position-wise.

Assuming, a pedestrian crosses the road only in an angle of 90° to the sensor, the approach gives a good estimation of the lateral velocity. In most of the test cases it outperforms the other candidates. Essential is the derivation of the ROI. If it is chosen too large the standard deviation grows. And if it is too small, the probability that it does not include the pedestrian rises. Currently, the approach does not regard standing pedestrians as they are not represented in accident statistics. But if the pedestrian starts moving, the tracker quickly detects a possible target.

  Figure 5: Tracked lateral position

Figure 5: Tracked lateral position

Tracked longitudinal velocity

Figure 6: Tracked longitudinal velocity

Test case with occultationFigure 7: Test case with occultation

6. Conclusion

Modern automotive safety systems have high requirements on the sensing devices, detecting dynamic objects in the vehicle’s environment to estimate a reliable collision risk for the desired actuation. A key factor in a predictive pedestrian safety system is a reliable tracking of objects likely causing a collision in a very short time horizon. Common radar systems rely on classical tracking methods like Kalman Filter that are based on several pre-processing stages e.g. threshold detection, clustering and extraction. In a highly frequent accident scenario where a pedestrian emerges out of occlusion these tracking methods typically fail as we have shown. The TBD approach described in this paper shows good estimation results of the velocities. These results were comparable with the data from the series radar or slightly better throughout all test scenarios. Moreover it is possible to use even weak reflections received by multipath propagations.

The particle filter was parameterized for detection of walking pedestrian, whereby a pedestrian could be separated from static objects like parked vehicles. Nevertheless, especially the estimation of the position has to be improved. At this stage, the very simple clustering of the particles leads to a high variation of the estimated position and thus causes and inacceptable update rate of the tracking.

In future work we will drive test scenarios with pedestrians standing on the curb and will find out how the TBF performs in multi-target situations with more than one pedestrian.



The authors would like to thank the European Commission for supporting this research work under the 7th Framework Programme (project ARTRAC, no. 284740).


[1] European Commision, “Road safety: Policy orientations on road safety 2011-2020,” Brussels, 2010.
[2] M.-M. Meinecke, M. Heuer, H. Rohling, S. Heuel et al., “User needs and Requirements for VRU protecting systems based on multipurpose narrow-band radar”, ARTRAC Deliverable D2.1, 2012.
[3] E. Coelingh, A. Eidehall and M. Bengtsson. “Collision Warning with Full Auto Brake and Pedestrian Detection - a practical example of Automatic Emergency Braking”. 13th International IEEE Conference on Intelligent Transportation Systems, 2010.
[4] M. Ooishi, “New Toyota Lexus Detects Pedestrians, Applies Brakes”, Tech-On!
[5] A. Bartsch, F.Fitzek and R.H. Rasshofer, “Pedestrian recognition using automotive radar sensors”, Adv. Radio Sci., 10, pp. 45-55, 2012.
[6] S. Heuel and H. Rohling, “Pedestrian Classification in Automotive Radar Systems,” 19th International Radar Symposium, pp.39-44, Poland, 2012.
[7] Y. Boers, H. Driessen, F. Gustafsson et al. “Track-before-detect algorithm for tracking extended targets,” IEE Proceedings - Radar, Sonar and Navigation, vol. 153, pp. 345-355, 2006.
[8] A. Broggi, P. Cerri, S. Ghidoni, P. Grisleri, H.-G. Jung, "A New Approach to Urban Pedestrian Detection for Automatic Braking", IEEE Transaction on ITS, vol. 10, pp. 594-605, 2009.
[9] V. C. Chen, F. Li, S.-S. Ho, and H. Wechsler, “Micro-Doppler Effect in Radar: Phenomenon, Model, and Simulation Study,” IEEE Transactions on Aerospace and Electronic Systems, vol. 42, no. 1, pp. 2–21, 2006.
[10] D. Sánchez, J. Häkli et al. ARTRAC - Deliverable D3.1: Architecture and Specification. 2013. [11] H. Rohling, “Some Radar Topics: Waveform Design, Range CFAR and Target Recognition” Advances in Sensing with Security Applications, Volume 2, pp. 293-322, 2006.
[12] M. Rutten, B. Ristic and N. Gordon, “A comparison of particle filters for recursive track-before-detect,” 7th International Conference on Information Fusion, pp. 169-175, 2005.
[13] Y. Boers and J. Driessen, “Particle filter based detection for tracking,” American Control Conference, pp. 4393-4397, 2001.
[14] S. Zecha, G. Juergens and P. Quittenbaum, “Inovative Test Methods and Facilities for Predictive Pedestrian Protection”, NH, 2013.

Michael Heuer Michael Heuer
Otto-von-Guericke-Universität Magdeburg
Ayoub Al-Hamadi Ayoub Al-Hamadi
Otto-von-Guericke-Universität Magdeburg
Marc-Michael Meinecke
Volkswagen AG


Get In Touch

  • Gelsingforsskaya 3, corpus 11D
    194044, St. Petersburg, Russia
  • +79522068225
  • This email address is being protected from spambots. You need JavaScript enabled to view it.