Expert Systems with Applications

Transcription

1 Expert Systems with Appications 37 (010) Contents ists avaiabe at ScienceDirect Expert Systems with Appications journa homepage: A neura network approach to target cassification for active safety system using microwave radar Seongkeun Park a, Jae Pi Hwang a, Euntai Kim a, *, Heejin Lee b, Ho Gi Jung c a Schoo of Eectrica and Eectronic Engineering, Yonsei University, C613, Sinchon-dong, Seodaemun-gu, Seou , Repubic of Korea b Department of Information and Contro Engineering, Hankyong Nationa University, Repubic of Korea c Mando Centra Research Center, Gyeonggi-do , Repubic of Korea artice info abstract Keywords: Target cassification Mutiayer perceptron neura network Active safety system Beief update Probabiistic cassification As a sensor in the active safety system of vehices, the microwave radar (MWR) woud be a good choice for the ocaization of the nearby targets but coud be a bad choice for their cassification or identification. In this paper, a target cassification system using a 4 GHz microwave radar sensor is proposed for the active safety system. The basic idea of this paper is that the pedestrians and the vehices have different refection characteristics for a microwave. A mutiayer perceptron (MLP) neura network is empoyed to cassify the targets and the probabiistic fusion is conduct over time to improve the cassification accuracy. Some experiments are performed to show the vaidity of the proposed system. Ó 009 Esevier Ltd. A rights reserved. 1. Introduction Safety has been a hot issue in recent vehicuar technoogy and a tremendous research has been conducted towards the direction. The researches concerning the safety of the passengers and drivers in the vehices have produced two paradigms: passive safety system and active safety system. The passive safety system purposes to minimize the damage after car accident and an air bag and safety bet beong to this cass (Chan, 007; Watanabe, Umezawa, & Abe, 1994). The passive system does not aim at reducing the possibiity of the car accidents. On the contrary, active safety system purposes to prevent the car accidents before they occur and it is now receiving much attention within vehicuar community. Mainy, the active safety system recognizes the surrounding environment around its own car and aerts the car driver about the nearby possibe dangers. The road sign recognition system (Nguwi & Kouzani, 006; Yoon, Lee, Kim, & Park, 008), and the bind spot warning system (Krips, Veten, Kummert, & Teuner, 004; US Patent , 005), the adaptive cruise contro (ACC) system (Bifuco, Simonei, & Di Pace, 008; Ioannou & Stefanovic, 005; Wang, hang, & Bubb, 007) and pedestrian protection systems (PPS) (Gandhi & Trivedi, 007) certainy beong to the active safety system. In the active safety system, the key technoogy is the understanding of the surrounding objects, that is, detection, tracking and identification of the nearby objects. For the purpose, severa * Corresponding author. E-mai address: etkim@yonsei.ac.kr (E. Kim). sensors are used and CCD cameras and range finders are the most common ones. The CCD camera returns rich information about the nearby target objects and provides reativey easy method for target recognition. However, it is difficut to measure the range to the nearby targets from the car. On the contrary, range finders such as a aser scanner or microwave radar easiy measure the ocation of nearby objects and are robust to the variation of the weather or time. But the range finders have difficuty in recognizing the target objects (Fuerstenberg & Dietmayer, 004; Fuerstenberg & Wihoeft, 001; Macachan & Mertz, 006). In this paper, we deveop a new microwave radar-based target cassification system for an active safety system. We do not use a CCD camera but use ony a 4 GHz microwave radar to cassify the nearby objects. Thus, with a singe radar, we can fufi both target tracking and target recognition simutaneousy at no extra cost. The basic idea of the target recognition by a microwave radar is that the pedestrians and the vehices have different refection characteristics for a microwave. Based on the idea, we buid a cassifier using a mutiayer perceptron neura network (MLP). Further, we present a probabiistic fusion method to make a cassification decision based on not ony the current measurement but aso a the past measurements. The remaining of this paper is organized as foows: in Section, the microwave radar sensor and the experimenta setup are expained. In Section 3, the probem is formuated and a neura network cassifier is designed for target cassification. In Section 4, the neura network cassifier resuts are fused over time to improve the cassification resut. Finay, some concusions are drawn in Section /$ - see front matter Ó 009 Esevier Ltd. A rights reserved. doi: /j.eswa

2 S. Park et a. / Expert Systems with Appications 37 (010) signa power and many features of a target is summarized in Radar Cross Section (RCS) equations (Knott, Shaeffer, & Tuey, 1993). Even though the reations expressed in equations are avaiabe, we do not use them directy since we cannot know the shape, size, or refectivity of the nearby targets. Instead, we gather some sampes from the radar and train the cassifiers to predict the objects. 3. Pedestrian and vehice cassification by mutiayer perceptron neura network 3.1. Mutiayer perceptron neura network cassifier Fig. 1. Configuration of active safety system.. 4 GHz microwave radar sensor In this paper, a 4 GHz microwave radar (MWR) named MAS- RAU005 is empoyed to sense and cassify targets around the vehice. The radar is manufactured by M/A-COM and is composed of a 4 GHz radar sensor and microprocessor unit (Manua of MA-COM MASRAU005, 005). Fig. 1 shows the configuration of our experimenta setup for target cassification. As in Fig. 1, the 4 GHz microwave sensor MASRAU005 is mounted on the car and is connected to a computer through CAN (Contro Area Network). In the system, we use a aptop computer equipped with P-IV and the program is run on Window XP. Fig. shows the experimenta setup in which two microwave radars are mounted on the vehice to be connected to a computer. The basic idea of this paper is to use the power of the microwave signa that is emitted from and returns to the radar, refected from the targets, and cassify the nearby objects. It means how much power of the microwave is attenuated in decibe when it is refected from the target. The received power is very sensitive and varies depending on fascia materia, paint, dirt and mud buidup, foam materia or other mechanica obstructions in the sensor fied of view. The reationship between the received Neura networks imitate the human brains and acquire knowedge about the hidden reationships between input and output directy from sampes (Duda, Hart, & Stork, 001; Hagan, Demuth, & Beae, 1995). Fig. 3 shows a three ayer MLP with input, hidden, and output ayers. Suppose we are given data set X ¼ fx 1 ; x ;...; x N gand each data x n ¼ x n 1 ; xn ;...; T xn m R m beongs to one of C casses x ð ¼ 1; ;...CÞ, where C is the number of casses, x k s are disjointed and X ¼ x 1 [ x [...[ x C. Then the output of neura networks shown in Fig. 3 is represented by ðx n Þ¼ 1 ðx nþ; ðx nþ;... C ðx T nþ ð1þ! X nh X m ðx n Þ¼f 1 w oh j f w hi ji xn i þ b jo þ b o!; ðþ j¼1 i¼1 where w hi ji and w oh j are weight vector between ith input node and jth hidden node, and between jth hidden node and th output node, respectivey. n h is the number of hidden node, f ðþis activation function. We train the MLP such that the output node associated with the true cass returns 1 and a remaining nodes return 0. Thus, we seect the target vaues by tðx n Þ¼½t 1 ; t ;...t C Š T t ðx n Þ¼ 1 if x n x 0 otherwise As in Duda et a. (001), we train the MLP such that the foowing function is minimized: ð3þ ð4þ Fig.. Experimenta setup in which 4 GHz microwave radar is mounted.

3 34 S. Park et a. / Expert Systems with Appications 37 (010) dx n þ pðx n ;x Þdx n ¼ XC pðx Þ pðxn ðx n Þ pðx n jx Þ Þdx n ¼1 þ pðx jx n Þpðx i jx n Þpðx n Þdx n ð6þ The second term of (6) does not invove W and thus we obtain ðx n Þpðx jx n Þ ð7þ JðWÞ ¼ XN n¼1 X C ¼1 ðx n Þ t ðx n Þ ¼ XC X X ðx n Þ 1 þ ¼1 x nx x nrx ¼ N XC N N 1 X N N ðx n Þ 1 þ N ¼1 x nx X! ðx n Þ 0 x nrx! ðx n Þ 0 N 1 N N where W is weight matrix and N is the number of data which beongs cass x.asnapproaches infinity, we appy Bayes rue, then 1 pðxn im pðx Þ ðx n Þ 1 jx Þdx n N!1 N JðWÞ~ JðWÞ¼ XC ¼1 þ pðx i Þ pðx Þ XC Fig. 3. Mutiayer perceptron with three ayers. ¼1 pðxn ðx n Þ 0 jx i Þdx n pðxn ðx n Þ Þdx n ð5þ ðx n Þpðx n ;x Þ That is, when the MLP is configured and trained as in (3) through (6), it returns the posterior probabiity of each cass. From the posteriors, we determine the cass by ¼ arg max ðx n Þarg max pðx jx n Þ: 3.. Target cassification using microwave radar Instead of using Radar Cross Section (RCS) equations (Knott et a., 1993), we coect the training data from the vehices and pedestrians and buid an MLP for cassification since the received power from the microwave depends on many factors of the target. We change the ocation of the target from 0m to 30 ms with an interva of 1 m, from 30 to +30 with the interva of 15 and measure the received power of the microwave for pedestrians and vehices. Fig. 4 shows the received power in case of zero bearing from both vehices and pedestrians. It can be observed that the received power is noisy and varies with respect to the distance between radar and objects. As it can be seen in the figure, there is a tendency that the vehices have arger received power than the pedestrians at the same distance and this is the key feature that we expoit in this paper. We coect 1853 vehice sampes and 0 pedestrian sampes and train an MLP as shown in Fig. 5. ð8þ Fig. 4. Received signa power (ero bearing).

4 S. Park et a. / Expert Systems with Appications 37 (010) (PÞ; X 1:t ¼fx 1 ; x ;...x t g denotes an accumuated sequence of reative distance, bearing and received powers from time 1 up to t. We name the new quantity as beief, and it is the accumuated posterior probabiity of x ¼ ( can be taken as a vehice (V) ora pedestrian (P)) conditioned on a current and past measurements. This beief measure is motivated by Thrun, Bugard, and Fox (005). As in Thrun et a. (005), we rewrite the beief of vehice cass into be t ðx ¼ VÞ ¼pðx ¼ VjX 1:t Þ ¼ pðx tjx ¼ V; X 1:t 1 Þpðx ¼ VjX 1:t 1 Þ pðx t jx 1:t 1 Þ ð11þ We use a three MLP with 5 nodes in the hidden ayer. We conduct sevenfod cross-vaidation and repeat it 10 times to test the statistica reiabiity. Tabe 1 shows the resut of 10 runs of sevenfod cross-vaidation. The microwave radar is noisy in nature, as shown in Fig. 4, and often returns outiers. The overa CCR (correct cassification rate) reaches 0.800, as shown in Tabe 1. The resuts are not bad but not satisfactory, either. In the next section, we improve the cassification rate by considering the probem in a probabiistic framework. 4. Probabiistic target cassification The target cassification based on MLP sometimes makes a wrong decision due to outiers produced by the microwave radar. To sove this probem, we utiize not ony the current measurement but aso a the previous measurements in a probabiistic framework to make a decision about the cass of the target. Here, the measurement means the features used in the MLP cassification: distance, bearing and received signa power. More specificay, in Section 3, we used the MLP ðx t Þpðx jx t Þ ð9þ in making a decision at time t and it actuay was the posterior probabiity based on the current measurement (Duda et a., 001). In this section, we define an accumuated posterior probabiity at time t be t ðxþ ¼pðxjX 1:t Þ Fig. 5. MLP cassifier. Tabe 1 Correct cassification rate (CCR) of 10 runs with sevenfod cross-vaidation using the MLP cassifier. Mean Variance Set Set Set Set Set Set Set Overa ð10þ based on not ony the current measurement x t but aso a the previous measurements X 1:t 1, where x is a binary variabe denoting whether the target is a vehice (VÞ or a pedestrian Since the current measurement x t is independent of X 1:t 1 conditioned on x, we can rewrite (11) into be t ðx ¼ VÞ ¼ pðx tjx ¼ VÞpðx ¼ VjX 1:t 1 Þ pðx t jx 1:t 1 Þ Pugging the Bayes rue, pðx t jx ¼ VÞ ¼ pðx ¼ Vjxt Þpðx t Þ pðx ¼ VÞ into (1) yieds be t ðx ¼ VÞ ¼ pðx ¼ Vjxt Þpðx t Þ pðx ¼ VjX1:t 1 Þ pðx ¼ VÞ pðx t jx 1:t 1 Þ In the same way, we compute the beief at time t for x ¼ P be t ðx ¼ PÞ ¼ pðx ¼ Pjxt Þpðx t Þ pðx ¼ PjX1:t 1 Þ pðx ¼ PÞ pðx t jx 1:t 1 Þ ð1þ ð13þ ð14þ ð15þ Then, dividing (14) by (15) and canceing some intractabe terms yieds be t ðx ¼ VÞ be t ðx ¼ PÞ ¼ pðx ¼ Vjxt Þ pðx ¼ Pjx t Þ pðx ¼ VjX1:t 1 Þ pðx ¼ PjX 1:t 1 Þ By pugging the outputs of the MLP pðx ¼ PÞ pðx ¼ VÞ ð16þ ðx t Þpðx jx t Þ ð17þ into (16), we obtain the foowing sequentia equation for the beiefs: be t ðx ¼ VÞ 1 be t ðx ¼ VÞ ¼ pðx ¼ Vjxt Þ 1 pðx ¼ Vjx t Þ pðx ¼ VjX 1:t 1 Þ ¼ VÞ 1 pðx 1 pðx ¼ VjX 1:t 1 Þ pðx ¼ VÞ ¼ gnn x¼v ðxt Þ 1 x¼v ðxt Þ be t 1ðx ¼ VÞ 1 be t 1 ðx ¼ VÞ 1 be 0ðx ¼ VÞ ¼ q be 0 ðx t V ¼ VÞ ð18þ Thus, from the previous beief be t 1 ðx ¼ VÞ (or be t 1 ðx ¼ PÞÞ and the current output of the MLP, we can update the current beief and we do not have to store a the measurement X 1:t 1 to compute the beief. Once q t V is obtained, be tðx ¼ Þ can be recovered by qt V be t ðx ¼ VÞ ¼ 1 þ q t V qt P be t ðx ¼ PÞ ¼ 1 þ q t P Then, we can make a cassification based on beief by x ¼ arg max½be t ðx ¼ ÞŠ ¼ arg max½pðx ¼ jx 1:t ÞŠ ¼P;V ¼P;V ð19þ ð0þ ð1þ

5 344 S. Park et a. / Expert Systems with Appications 37 (010) Fig. 6. The proposed probabiistic cassification strategy for the active safety system. The probabiistic cassification strategy is summarized in Fig. 6. In summary, the MLP makes soft decision on the target based on the current radar outputs and the soft decision of the MLP is fused over time in the Bayesian probabiistic framework as shown in Fig Experimenta resuts In this section, we conduct some experiments to show the effectiveness of the proposed method. In the probabiistic cassification, we set both be 0 ðx ¼ VÞ and be 0 ðx ¼ PÞ to 0.5 since we have no prior information about the target. We use the same MLP trained from 1853 vehice sampes and 0 pedestrian sampes in the previous section. Fig. 7 shows our experimenta setup in which both the vehice and the pedestrian move at amost the same speed. MLP makes a soft decision and it is fused over time when the targets are moving. Fig. 8a and b shows specific exampes of probabiistic cassification resut based on the MLP for a vehice and a pedestrian, respectivey. As can be seen in the figures, the MLP often makes wrong decisions because the microwave radar is noisy but the probabiistic fusion compensates misbehavior of the MLP successfuy by taking into account the previous measurements in the probabiistic framework. As in Section 3, we conduct sevenfod cross-vaidation and repeat it 10 times to test the statistica reiabiity. Tabe shows the resut of the probabiistic cassification for sevenfod cross-vaidation. The overa CCR reaches and is improved from the MLP by amost Concusion In this paper, we have deveoped a target cassification for active safety system. We impemented it using ony using a microwave radar without the hep of any CCD camera. Since we used ony a 4 GHz radar, we can fufi both target tracking and recognition simutaneousy at no extra cost. Our system was buit and its vaidity was demonstrated through a rea word experimentation. The contribution of this paper is threefod: (1) Target cassification system for an active safety system is impemented with ony microwave radar and a power refected from the target is used as key feature. () An MLP neura network is trained to cassify the target based on the radar outputs. (3) The cassification performance is greaty improved by combining the cassification resuts over time in the probabiistic framework. Fig. 7. A pedestrian and a vehice moving at the same speed. However, more works sti shoud be done such that our system woud be more stabiized and be instaed in commercia vehices.

6 S. Park et a. / Expert Systems with Appications 37 (010) Fig. 8. Comparison of MLP cassifier and probabiistic cassifier. Acknowedgements This work is supported by Mando Co., Active Protection Pedestrian System Project. References Bifuco, G. N., Simonei, F., & Di Pace, R. (008). Experiments toward an human-ike Adaptive Cruise Contro. In Inteigent vehices symposium (pp ).

7 346 S. Park et a. / Expert Systems with Appications 37 (010) Tabe CCR of 10 runs of sevenfod cross-vaidation using probabiistic cassifier. Mean Variance Set Set Set Set Set Set Set Overa Chan, C. Y. (007). Trends in crash detection and occupant restraint technoogy. IEEE Proceedings, 95(), Duda, R., Hart, P., & Stork, D. (001). Pattern cassification. Addison Wesey. Fuerstenberg, K. & Dietmayer, K. (004). Object tracking and cassification for mutipe active safety and comfort appications using a mutiayer aser scanner. In IEEE inteigence vehice symposium (pp ). Parma, Itay. Fuerstenberg, K., & Wihoeft, V. (001). Object tracking and cassification using aserscanners Pedestrian recognition in urban environment. In Inteigence transportation systems conference (pp ). Oakand, USA. Gandhi, T., & Trivedi, M. M. (007). Pedestrian protection systems: Issues, survey, and chaenges. IEEE Tranactions on Inteigence Transportation Systems, 8(3), Hagan, M. T., Demuth, H. B., & Beae, H. (1995). Neura network design. USA: PWS Pubishing Company. Ioannou, P. A., & Stefanovic, M. (005). Evauation of ACC vehices in mixed traffic: Lane change effects and sensitivity anaysis. IEEE Transactions on Inteigence Transportation Systems, 6(1), Knott, E. F., Shaeffer, J. F., & Tuey, M. T. (1993). Radar cross section. Boston: Artech House. Krips, M., Veten, J., Kummert, A., & Teuner, A. (004). AdTM tracking for bind spot coision avoidance. In Inteigent vehices symposium, 004 IEEE (pp ). Macachan, R., & Mertz, C. (006). Tracking of moving objects from a moving vehice using a scanning aser rangefinder. In Inteigence transportation systems conference (pp ), Toronto, Canada. Manua of MA-COM MASRAU005, 005. Nguwi, Y.-Y., & Kouzani, A.. (006). A study on automatic recognition of road signs. In 006 IEEE conference on cybernetics and inteigent systems (pp. 1 6). Thrun, S., Bugard, W., & Fox, D. (005). Probabiistic robotics. Cambridge, MA: MIT Press. US Patent (005). Bind spot warning system for an automotive vehice. US Patent Issued on February, 005. Wang, W., hang, W., & Bubb, H. (007). Car-foowing safety agorithms based on adaptive cruise contro strategies. In Inteigent systems and informatics, 007. SISY 007. Fifth internationa symposium (pp ). Watanabe, K., Umezawa, Y., & Abe, K. (1994). Advanced passive safety system via prediction and sensor fusion. In Vehice navigation and information systems conference (pp ). Yoon, C., Lee, H., Kim, E., & Park, M. (008). Rea-time road sign detection using fuzzy-boosting. IEICE Transactions on Fundamentas of Eectronics Communications and Computer Sciences, E91-A(11),