DRIVER MODELS TO I CREASE THE POTE TIAL OF AUTOMOTIVE ACTIVE SAFETY FU CTIO S

18th Europen Signl Processing Conference (EUSIPCO-2010) Alborg, Denmrk, August 23-27, 2010 DRIVER MODELS TO I CREASE THE POTE TIAL OF AUTOMOTIVE ACTIVE SAFETY FU CTIO S J. Sjöberg, E. Coelingh*, M. Ali*, M. Brännström*, nd P. Flcone Deprtment of Signls & Systems, Chlmers University University, SE41296 Gothenburg, Sweden * VOLVO CAR CORPORATIO, DEPT. ACTIVE SAFETY FU CTIO S, GOTHE BURG, SWEDE PHO E: + (46) 31-772 1855, FAX: + (46) 31-772 1782, EMAIL: JO AS.SJOBERG@CHALMERS.SE WEB: WWW.CHALMERS.SE/S2 ABSTRACT This pper describes how the potentil of some utomotive ctive sfety functions depend on the used driver model. It is shown tht by including more dvnced driver model, it is possible to enhnce the use of the signls from different sensor systems to let the ctive sfety function intervene erlier nd smoother so tht the drivers re disturbed less, nd the chnce to void n ccident increses. 1. I TRODUCTIO Automotive ctive sfety functions wrn or intervene in criticl situtions to help the driver void or mitigte ccidents. During norml driving, when the driver hs the sitution under control, the ctive sfety functions re t rest. A function with intrusive interventions cn be irritting nd disturbing if it intervenes when the driver hs the sitution under control. Hence, such events must be voided nd this is the motivtion to the prdigm to llow n intervention only if drivers cnnot void the ccident themselves. The tsk to decide if the driver needs ssistnce is clled thret ssessment, see e.g., [1], [2], nd the prdigm indictes tht knowledge of the cpbility of the driver is essentil for good thret ssessment lgorithm. The thret ssessment module is signl processing unit where the trffic sitution is ssessed by using e.g. vehicle models nd driver models in combintion with signls tht describes the current stte of the vehicle nd the surrounding objects. For exmple, effective collision voidnce technology is bsed on the bility to predict the ner future. In timefrme of less thn 1 second, the future cn be predicted well using vehicle kinemtics or vehicle dynmics models, but t lrger prediction horizons, the driver behviour strts to ply dominnt role. E.g. the risk for collision with n object 400 m hed of the vehicle is lmost solely dependent on the driver behviour. When one cn predict the driver behviour one could potentilly judge whether the driver intends to brke, steer or ccelerte to void collision or not. In cse it is judged tht the driver hs no such intentions, wrning or utomtic interventions could be pplied much erlier, thereby improving the benefit of the intervention without incresing the risk of disturbing the driver with unnecessry interventions. The discussion nd the conclusions in this pper build on three exmples where successively more dvnced driver models re being used. The first exmple considers utonomous brking in rer-end collisions nd no driver model is included. Insted it is ssumed tht mximum brking nd steering cn be relized immeditely. With this description, there is theoreticl possibility to void collision very lte by steering nd, hence, the utonomous brking cn strt only when this theoreticl chnce hs diminished. Nevertheless, such system cn mitigte rer-end collision so tht the impct speed is reduced. This exmple is described in Section 3.1. The second exmple considers the rection time of the driver nd how fst the driver cn move the steering wheel. Including these fetures in the description of the driver, the possibility to steer wy from possible rer-end collision decreses compred to the cse with immedite control ctions. Hence, the ctive sfety system cn intervene with utonomous brking erlier which gives improved chnces to void n ccident, s well s lrger reductions of impct speed. This exmple is described in Section 3.2. In the third exmple the risk of off-rod ccidents in curves due to too high speed is considered. The driver is described s controller with the objective to follow the rod. Using preview informtion bout the rod, nd mthemticl models of the vehicle nd the driver, the driving hed of the present position of the vehicle cn be simulted. In this wy it is possible to predict criticl situtions where the rod friction does not suffice to give the tire forces to hold the vehicle on the rod. This mkes it possible to issue n erly wrning, or to utomticlly reduce the velocity in gentle wy, before the criticl sitution occurs. This is in contrst to, e.g. n ESP system which is ctivted once the loss of control is imminent. The dvntge is tht one potentilly cn void more ccidents with less intrusive intervention. This exmple is described in Section 3.3. 2. SYSTEM ARCHITECTURE In Figure 1 n overview of generl ctive sfety system is depicted. The simpler ctive sfety functions only EURASIP, 2010 ISSN 2076-1465 204

mke use prts of the rchitecture. For exmple, for nti-lock brkes (ABS) nd yw-stbility control systems (ESP) the Tret Assessment, Decision Mking nd the Interven- stte giving tion Module only use signls bout the vehicle informtion bout if wheel is bout to slide. This mens tht no informtion bout the driver or the environment is used. More dvnced systems such s lne deprture wrning systems, nd rer-end collision mitigtion by brking lso mke use of signls giving informtion bout the vehicle surroundings, the Environment Informtion block in Figure 1 s input to the Tret Assessment nd Intervention Module. Hence, such systems require sensors nd signl processing delivering this informtion. The next step, which is the focus of this pper, is the im- the provement which is possible to obtin by lso considering driver s prt of the system. This mens thtt the thret stht it cn re- sessment block contins driver model, nd ceive signls from the driver s ctions illustrted with the Driver block in Figure 1. With driver model, in combin- tion with informtion bout the environment, it is possible to mke predictions with lrger horizon on which new ctive sfety systems cn be bsed. The potentil of these systems re of course dependent on the qulity of the driver models nd the sensor systems. Figure 1. Possible generl description of n ctive sfety system. Simple thret ssessment lgorithms, like those for ABS nd ESP, mke use of informtion only bout the vehicle stte. Next step is to include informtion from the environment like in some lne keeping systems. More dvnced lgorithms lso mke use of informtion bout the driver. 3. THREAT ASSESSME T EXAMPLES 3.1 Thret Assessment without Driver Model In its simplest form thret ssessment does not tke driver behviour into ccount. The first genertion Volvo's of Colon thret s- lision Wrning with Auto Brke [3] is bsed sessment tht mkes use of velocity mesurements nd constnt ccelertion model for the behviour of the host nd led trget vehicle. The motion of the host vehicle in reltion to the led trget vehicles is nlyzed nd possible collision event is sid to be imminent s soon s neither the mximum steering nor mximum brke ction could led to n voidnce of the rer-end collision. In terms of cceler- tions this mens tht s soon s the mximum chievble lterl nd longitudinl ccelertion due to steering nd brking ction is less thn the needed respective cceler- of the needed ccel- tions, collision is imminent. The rtio ertion nd mximum chievble ccelertion for brking nd steering ctions re denoted s brking thret number (BTN) nd steering thret number (STN), respectively, in the following wy, BT x, needed = nd ST = x, vilble y, needed y, vilble In [1] BTN nd STN re described s quntifiers for the collision risk. The vilble ccelertions re dependent on the tire-to-rod friction. When both BTN nd STN re lrger thn 1 collision is judged to be imminent, i.e. physiclly unvoidble, nd utomtic emergency brking is pplied. As depicted in Figure 2, n ccident cnnot be voided for velocities lrger thn 25km/h. For lrger speeds the impct speed will however be reduced, nd thus the consequences for the occupnts [4]. Using this criterion gives low risk for flse interven- norml driving, but tions tht could disturb the driver during the benefit of the system would improve drsticlly if the impct speed ws reduced even more due to n erlier uto brke intervention. One wy of chieving this is to incorporte ctutor dy- tht longitudinl nmics. In the model bove it is ssumed nd lterl ccelertion cn be chieved immeditely. But due to ctutor nd vehicle dynmics this is not possible. Knowing tht it tkes time to build up this ccelertion one could dpt the BTN nd STN clcultions, s suggested in [1]. This will led to erlier intervention nd thus incresed benefit. 3.2 Thret Assessment with Simple Driver Model Insted of strting n utomtic intervention when it is physiclly impossible to void collision, one could strt utomtic intervention when it is judged impossible for the driver to void collision. Using the ltter requires tht one incorportes driver model into the thret ssessment model together with the velocity signls. One reltive simple wy of doing this is by judging whether the driver is distrcted. In [5] sitution ssessment lgorithm is proposed tht estimtes whether the driver of the host vehicle is distrcted or not. The lgorithm uses the fct tht ccident reserch shows tht up to 93% of ll drivers were distrcted just before the rer-end collision occurred [6] suggesting tht the driver did not steer due to the distrction. The lgorithm strts to clssify the driver s either ctive or pssive. In the second step, only pssive drivers cn be clssified s distrcted. The driver is clssified s ctive when ω > 1 rd/s (1) 205

or ω > 0.5 rd/s continuously for t lest 0.5s during the preceding 1.0s, where ω is the steering wheel ngle chnge rte. The rest of the time, the driver is ssessed s pssive. Secondly, the lgorithm ssesses possible threts from led vehicles in terms of the STN. The driver is now ssessed s distrcted if STN exceeds 0.5 nd the driver is ssessed s pssive nd the led vehicle hs velocity below 10km/h. The dditionl impct speed reduction tht cn be chieved when the driver is correctly ssessed s distrcted is depicted in Figure 2. 2) Similrly s for steering mnoeuvres, the prmeteriztion of potentil brking nd ccelerting mnoeuvres is selected s constnt jerk up to mximum ccelertion. This profile corresponds well to ccelertion profiles for mny brke ctutors with limited ccelertion chnge rte nd limited decelertion cpcity. Dynmic effects of these mnoeuvres were lso included in the vehicle model, for exmple dely in the steering. The result of incorporting this pproch for rer-end collisions is depicted in Figure 3. By including model of the dynmics of the driver, mnoeuvre voiding collision needs to be initited erlier. Hence, the utomtic brking cn intervene erlier which gives lrger velocity reduction. Figure 2. Reltive velocity reduction for different host vehicle speeds when driving stright towrds sttionry led vehicle, showing tht incorporting driver distrction (ST = 0.5) increses potentil s compred with not incorporting this (ST = 1). Additionlly, performing erlier interventions when the driver is ssessed s being distrcted do not cuse ny dditionl flse interventions s verified in set of 200 driving hours in rel trffic conditions. This illustrtes how even simple driver model cn improve the benefit of n ctive sfety feture. Another exmple of reltive simple driver model is provided in [7]. A driver does not chieve mximum ccelertion immeditely when brking or steering vehicle. So the potentil voidnce mnoeuvres re prmeterized such tht they represent common driver behviour. One prmeteriztion is presented for steering mnoeuvres nd nother for brking or ccelertion mnoeuvres. Ech type of mnoeuvre is described using only one prmeter, which is selected such tht the severity of the mnoeuvre increses with n incresing prmeter vlue. The vehicle speed is ssumed to be constnt while the driver steers, v = v 0, nd the curvture is ssumed to be constnt while the driver brkes or ccelertes, c = c0. 1) Steering mnoeuvres: During norml driving, drivers often use steering mnoeuvres with constnt steering wheel ngle rte followed by constnt steering wheel ngle [8]. So nturl prmeteriztion for steering mnoeuvres is to hold the steering wheel ngle rte constnt during limited time intervl, until reching finl steering wheel ngle. Figure 3. Velocity reduction versus host vehicle speeds when driving stright towrds sttionry led vehicle. The dshdotted line shows the result using constnt ccelertion model (no driver model). The dshed line shows the performnce when the driver model is included. Also this lgorithm ws verified on dt set of 200 hours of driving without ny flse interventions. This gin shows tht incorporting knowledge bout the driver, in combintion with signls bout the trffic sitution, cn increse the benefit of n ctive sfety function without significntly incresing the risk for flse interventions. 3.3 More Advnced Driver Model The next step in the direction of more dvnced driver models is to consider the dynmics of the closed loop system of driver nd vehicle. With more dvnced informtion system giving informtion bout the curvture of the rod hed of the vehicle, the objective of the driver is to follow the rod. In this context, the driver cn be seen s the controller mnging the vehicle to follow the rod. In this wy n ctive sfety function cn be defined, bsed on vehicle dynmic models nd driver models in connection with dvnced sensing technologies providing informtion bout rod geometry, globl position, velocity, nd possibly, moving objects. This technique is introduced in [9] for rodwy deprture prevention system nd further developed in [10]. An illustrtion of this is shown in Figure 4 where the vehicle is pproching curve nd the lgorithm clcultes prediction of the tire slip ngles in the curve by simulting the closed-loop system of driver nd vehicle using sensor informtion bout position, velocity nd rod curvture s inputs signls. 206

Figure 4. The vehicle is pproching curve nd the lgorithm predicts the vehicle trjectory time horizon hed of the vehi- wrn- cle nd clcultes the necessry tire forces. If the prediction indictes too lrge tire slip ngles, ing cn be issued, or n intervention cn be issued so tht the velocity is decresed. Since this cn be done slightly before the criticl sitution, less intrusive ction my suf- issue. fice compred to wht n ESP system would need The driver model used for the steering in [9] is described by (2) where is the steering ngle which depends on the signls, the vehicle's lterl displcement from the rod t dis- nd, the tnce ds hed of the vehicle's centre of grvity, difference the vehicle's heding ngle nd the rod s hedseconds hed ing ngle t the preview point which is of the vehicle. Figure 5 illustrtes the driver model signls. Figure 6. Mesured steering ngle together with the steering ngle given by the model (5). A) Estimtion dt, B) Vlidtion dt. It is simple driver model. The first term gives feedbck contribution if the vehicle hs n offset in the lne nd the second term is feed-forwrd forwrd contribution which depends on the curvture hed of the vehicle. Using dt from test drive, the proposed lgorithm is vlidted ginst the existing ESP. The result is depicted in Figure 7. It shows the predicted tire slip ngle together with n indicting function showing when the ESP system ctivtes. From the figure it is cler tht the lgorithm indictes lrge slip ngles pproximtely 1-2 seconds before the ESP system ctully ctivtes. Figure 5. Illustrtion of the mesures used model (2). The prmeters in the driver model,, estimted using mesurements. See Figure 6. in the driver nd re Figure 7. Solid: predicted d tire slip ngle, dshed: function indicting when the ESP system is ctivted. High slip ngles is cler indictor of imminent criticl sitution. Hence, this time window hs become vilble thnks to the proposed lgorithm nd it cn be used for providing n erlier wrning or intervention in order to minimize driver disturbnce. Or, it cn be used to strt interventions erlier in order to increse the sfety benefit. For under-steered vehicles the preview of n imminent loss of control with n ctivtion tion of the ESP system cn be 207

significnt for voiding loss of control. The reson for this is tht the vehicle s verticl lod distribution in curve reduces the norml force t tht specific wheel where the ESP system needs to brke to stbilize the vehicle. Hence, the ESP system might hve lost most of its power to stbilize the vehicle. This is illustrted in Figure 8 nd further described in [11]. Figure 8. Illustrtion of vehicle s verticl lod distribution in curve sitution. The ellipses represent vilble friction t ech wheel, the sizes of the ellipses depend on the verticl lod nd the forces produced t the tires re constrined to lie within the ellipses. In curve sitution, much of the vehicle s weight is redistributed to the outer side. As cn be seen vilble brke force t the inner bck wheel is gretly reduced. 4. DISCUSSIO A D CO CLUSIO S Using three exmples it hs been shown tht knowledge on driver behviour cn be used to enhnce ctive sfety functionlity. Together with sensor systems giving informtion bout the trffic sitution, driver model mkes it possible to mke predictions bout trffic sitution over longer time horizon. This cn be used for providing erlier nd more comfortble interventions or for incresing the sfety benefit, without significntly incresing the risk of flse or unnecessry interventions. The key for chieving this potentil benefit is n ccurte driver model. Typicl chllenges in obtining driver model re: Understnding driver behviour in ner-criticl situtions. Deling with driver vribility, s driver behviour chnges over time nd vries between individuls. Obtining on-line dt on driver behviour. This cnnot be mesured directly nd hs to be estimted from e.g. vehicle behviour nd driver inputs trough pedls nd steering wheel. These re topics for future reserch nd the result will influence our dily life in trffic. REFERE CES [1] J. Jnsson, Collision Avoidnce Theory with Appliction to Automotive Collision Mitigtion. PhD. Thesis, Linköping University, 2005, Linköping Studies in Science nd Technology. Disserttions. No. 950. [2] J. Hillenbrnd, A.M. Spieker, nd K. Kroschel, ""A multilevel collision mitigtion pproch - Its sitution ssessment, decision mking nd performnce trdeoffs," IEEE Trns. Intell. Trnsp. Syst., vol. 7, no. 4, pp. 528-540, 2006. [3] E. Coelingh, H. Lind, W. Birk, nd D. Wetterberg, "Collision Wrning with Auto Brke," in FISITA World Congress, Yokohm, 2006. [4] E. Coelingh, L. Jkobsson, H. Lind, nd M. Lindmn, "Collision Wrning with Auto Brke A Rel-Life Sfety Perspective," in 20th Int. ESV Conf., 2007. [5] M. Brännström, E. Coelingh, nd J. Sjöberg, "A sitution nd thret ssessment lgorithm for rer-end collision voidnce system," in IEEE Intelligent Vehicles Symposium, 2008, pp. 102-107. [6] V.L. Nele, T.A. Dingus, S.G. Kluer, J. Sudweeks, nd M. Goodmn, "An overview of the 100 cr nturlistic study nd findings," in 19th ESV Conf, 2005. [7] M. Brännström, E. Coelingh, nd J. Sjöberg, "Model- Bsed Thret Assessment for Avoiding Arbitrry Vehicle Collisions," IEEE Trnsctions on Intelligent Trnsporttion Systems, 2010, Accepted for publiction. [8] H. Godthelp, "Vehicle Control During Curve Driving," Humn Fctors, vol. 28, no. 2, pp. 211-221, 1986. [9] M. Ali, P. Flcone, nd J. Sjöberg, "A Predictive Approch to Rodwy Deprture Prevention," in 21st IAVSD Symposium on Dynmics of Vehicles on Rods nd Trcks, Stockholm, 2009. [10] P. Flcone, M. Ali, nd J. Sjöberg, "Set-Bsed Thret Assessment in Lne Guidnce Applictions," in IFAC Symposium Advnces in Automotive Control, Munich, Germny, 2010. [11] M. Ali, C. Olsson, nd J. Sjöberg, "Towrds Predictive Yw Stbility Control," in IEEE Intelligent Vehicles Symposium, 2009. 208