**3.3. Classification based on the tracker feedback**

A single radar measurement and multiple radar measurements were considered for feature extraction. These features are already a good basis for the classification process. In this subchapter, a third, additional feature set is described. This set can be extracted from the tracking algorithm. In this adaptive algorithm, different process noise of pedestrian radar measurements and vehicle radar measurements result in different gains for the track. On this basis, the process noise *Q* and the calculated gain *K* are additional features and are added to single and multiple radar measurement features in the classification process.

### *3.3.1. Feature extraction*

Tracking is defined by a state estimation of moving targets. This state estimation is determined from the state parameters such as position, velocity and acceleration from a detected target. Known tracking methods are for example the alpha-beta filter or Kalman filter [26] which estimate a new state using a well-known prior state (e.g. position, velocity, acceleration). This reduces false alarms and smoothes movements of the objects.

12 Will-be-set-by-IN-TECH

0

Vehicle Pedestrian Other

0.2

0.4

0.6

Probability

**Figure 9.** Feature histogram of the velocity profile using multiple radar measurements as a basis for

Additionally, classifying feature vectors from single and multiple measurements results in

Vehicle 90.71 0.58 8.71 Pedestrian 4.94 53.94 41.12 Other 16.28 16.64 67.08

**Table 5.** Confusion matrix: classification applied to a single and multiple measurement test data set

Due to additionally multiple measurements as a basis for feature extraction, an improvement in the classification result is shown. Especially in terms of correct classification of a pedestrian and false alarms in which a pedestrian was classified as a vehicle, a significant enhancement

A single radar measurement and multiple radar measurements were considered for feature extraction. These features are already a good basis for the classification process. In this subchapter, a third, additional feature set is described. This set can be extracted from the tracking algorithm. In this adaptive algorithm, different process noise of pedestrian radar measurements and vehicle radar measurements result in different gains for the track. On this basis, the process noise *Q* and the calculated gain *K* are additional features and are added to

Tracking is defined by a state estimation of moving targets. This state estimation is determined from the state parameters such as position, velocity and acceleration from a detected target. Known tracking methods are for example the alpha-beta filter or Kalman filter [26] which

single and multiple radar measurement features in the classification process.

0.8

1

0 1 2 3 4 5

(b) *v*profile,buf, Vehicle

v profile [m/s]

0 1 2 3 4 5

feature extraction. The common area index is calculated to 0.88.

**3.3. Classification based on the tracker feedback**

(a) *v*profile,buf, Pedestrian

v profile [m/s]

0

fewer false positives.

containing 8000 data samples.

*3.3.1. Feature extraction*

is seen.

0.2

0.4

0.6

Probability

0.8

1

The Kalman filter is a linear, recursive filter, whose goal is to determine an optimal estimate of the state parameters. The optimal estimate is based on available measurements and the models which describe the observed objects. In the equation of the motion model and observation model, the measurement noise is assumed to be average free, white Gaussian noise with the known covariance *Qk*−<sup>1</sup> and *Rk* respectively. Under the given conditions, i.e., linear models and Gaussian statistics, the Kalman filter provides the optimal solution for the estimation of the state in the sense of minimizing the mean squared error, as described in [26].

The tracking for the object described by the motion model of the Kalman filter works fine as long as the motion models fit to the object. Pedestrians, vehicles, and static objects have different motion, which makes the tracking more difficult. Instead of creating a different motion and observation model for each object, it is proposed to determine the covariance of process noise *Qk* and the measurement noise *Rk* adaptively. The process noise considers a non-modeled behavior in the motion model, while the measurement noise consideres uncertainty in the measurement. The original Kalman filter is not adaptive, which is why deviations from the model can not be handled. The gain matrix *K*, which is calculated from the process and measurement noise, reaches a stable condition after a short measurement time. An increase in the covariance *Qk* leads to a larger value for *K*, so that the measured values are weighted more strongly, a decrease in *Qk* relies more on the estimation.

In addition to an improved tracking effect, additional features can be extracted from the adaptive adjustment of the process noise, as pedestrian measurements in range and velocity differ from those of vehicles. Next to the process noise *Qk*,*<sup>v</sup>* of the velocity, the Kalman gain *Kv* (velocity component of the matrix *K*) is a good feature as measurements show. Anyhow, in an adaptive adjustment of the process noise, a compromise between the compensation of non-modeled movements and the filtering effect to reduce noise must be found, even though features are extracted.

The process noise matrix *Q* describes object-specific measurement properties that are initially set and are readjusted during operation of the tracker. For example, the readjustment of a single coefficient *Qk*,*<sup>v</sup>* in the *Q* matrix at the measurement *k* in respect to the velocity *v* is based on the actual target range *R*, the velocity *v*, the parameters *a* and *b* in an alpha-beta filter. Equation (11) shows the relation. The velocity of a pedestrian deviates between consecutive measurements, while the velocity deviation of a car within consecutive measurements is small or even zero. Consequently, the velocity *v* can be used to update the matrix *Q*.

$$Q\_{k, \upsilon} = (1 - \beta) Q\_{k-1, \upsilon} + \beta \frac{|z\_{k, \upsilon} - v\_{k|k-1}|}{R\_{k|k} \cdot a + b} \tag{11}$$

The predicted covariance matrix *Pk*|*k*−<sup>1</sup> in the tracking process depends on the motion model *Fk*−1, the currently measured covariance *Pk*−1|*k*−<sup>1</sup> and the process noise *Qk*−<sup>1</sup> as shown in Equation (12).

$$P\_{k|k-1} = F\_{k-1} P\_{k-1|k-1} F\_{k-1}^T Q\_{k-1} \tag{12}$$

#### 14 Will-be-set-by-IN-TECH 254 Ultra-Wideband Radio Technologies for Communications, Localization and Sensor Applications Pedestrian Recognition Based on 24 GHz Radar Sensors <sup>15</sup>

Under the use of this covariance matrix *Pk*|*k*−<sup>1</sup> and the innovation covariance *Sk*, the gain *Kk* can be calculated.

$$K\_k = P\_{k|k-1} S\_k^{-1} \tag{13}$$

Vehicle Pedestrian Other

Pedestrian Recognition Based on 24 GHz Radar Sensors 255

Vehicle 92.84 0.50 6.66 Pedestrian 5.71 61.22 33.07 Other 10.06 13.30 76.64

**Table 7.** Confusion matrix: classification applied to a single, multiple and tracker feedback measurement

This chapter described a pedestrian classification algorithm for automotive applications using an automotive 24 GHz radar sensor with a bandwidth of 150 MHz as a measuring device. Three different systems for pedestrian recognition have been considered. The first system was based on a single radar measurement. The second system extracted a feature set on the basis of multiple radar measurements. Finally a tracking procedure was adapted to extract an additional feature set. The results show an increasing performance in the classification accuracy by using single-, multiple- and tracker feedback features. It is also pointed out that is not necessary to equip radar sensors with large bandwidths in order to classify pedestrians

*Department of Telecommunications, Hamburg University of Technology, Hamburg, Germany*

[1] Foelster, F.; Oprisan, D.; Rohling, H., (2004) Detection and Tracking of extended targets for a 24GHz automotive radar network, International Radar Symposium IRS 2004,

[2] Meinecke, M.-M.; Obojski, M.; Toens, M.; et.al, (2003) Approach For Protection Of Vulnerable Road Users Using Sensor Fusion Techniques, International Radar

[3] Schiementz, M.; Foelster, F. (2003) Angle Estimation Techniques for different 24 GHz Radar Networks, International Radar Symposium IRS 2003, Dresden, Germany, 2003 [4] Rohling, H.; Heuel, S.; Ritter, H. (2010) Pedestrian Detection Procedure integrated into

[5] Schmid, V.; Lauer, W.; Rollmann, G. (2003) Ultra Wide Band 24 GHz Radar Sensors for Innovative Automotive Applications, International Radar Symposium IRS 2003,

[6] Oprisan, D.; Rohling, H.,(2002) Tracking Systems for Automotive Radar Networks, IEE

[7] Philomin, V.; Duraiswani, R.; Davis, L. (2000) Pedestrian tracking from a moving vehicle, Proceedings of the IEEE Intelligent Vehicle Symposium 2000, Dearborn (MI), USA [8] Ran, Y. et al (2005) Pedestrian classification from moving platforms using cyclic motion pattern, IEEE International Conference on Image Processing 2005, Volume 2, pages: II

an 24 GHz Automotive Radar, IEEE RADAR 2010, Washington D.C.

test data set containing 8000 data samples. **4. Summary and conclusions**

Hermann Rohling and Steffen Heuel

Warsaw, Poland; pp 75-80

Dresden, Germany, pp 119-124

Radar 2002, Edinburgh

854-7

Symposium IRS 2003, Dresden, Germany, pp 125-130

in urban areas.

**Author details**

**5. References**

Both, matrix *Qk* and gain *Kk* are used as additional features in the classification process. The separation of the feature space is depicted in Figure 10. These histograms show that a vehicle (Figure 10(c), 10(d)) has lower process noise *Qv* and thus results in a smaller gain *K* compared to a pedestrian (Figure 10(a), 10(b)). This is due to a mostly linear trajectory of a vehicle with one main reflection point. A pedestrian echo fluctuates, which is reflected in a greater process noise and thus larger gain *K*.

**Figure 10.** Feature Histogram of Process Noise *Qv* and Gain *K*.

The additional features *Q* and *K* are available for a created and active track. For each detected object a track is created, which is however not activated until several measurements can be associated with the track. A track must be confirmed by subsequent measurements, otherwise the track remains semi-active. In any case, additional features are available for classification.


**Table 6.** Additional Feature Set extracted from the tracker using an adaptive process noise.

### *3.3.2. Classification*

An additional feature extraction based on the process noise and Kalman gain has been described. These features are added to the prior feature set. Table 7 shows the results of the classifier using all proposed features, single measurement, multiple measurements and tracker feedback. Even in an urban area with a high density of static targets pedestrians could be detected, tracked and classified with a true positive rate of 61.22% in the test data set. These results outperform prior outcomes using single- and the multiple measurements as a feature basis.

The classification results show an increasing performance in terms of correct classification and misclassification of pedestrians.


**Table 7.** Confusion matrix: classification applied to a single, multiple and tracker feedback measurement test data set containing 8000 data samples.
