IROS2019国际学术会议论文集 1415

On Modeling the Effects of Auditory Annoyance on Driving Style and Passenger Comfort Edson Araujo1 Michal Gregor2 Isabella Huang3 Erickson R Nascimento1 and Ruzena Bajcsy3 Abstract Despite the impressive progress being made in autonomous vehicles human drivers will remain ubiquitous in the imminent years Therefore intelligent hybrid vehicular systems must be aware of the interactions between humans and the environment e g sound vibration speed etc In this paper we evaluate the effect of acoustic annoyance on drivers in a real world driving study We found signifi cant differences in driving styles elicited by annoying acoustics and present an online classifi er that uses onboard inertial measurement unit measurements to distinguish whether a driver is annoyed with 77 accuracy Moreover we directly measured the forces applied on the passenger with a pressure mat lined on the car seat and empirically confi rm that our proposed passenger dynamics model is reasonable However due to our acoustically induced driving styles not being polarizing enough we were unable to show that passengers self reported ride comfort changed with acoustic annoyance I INTRODUCTION Rapid progress is being made in intelligent vehicular systems as developmental milestones in perception control and navigation are compounding Despite the importance of developing algorithms for autonomous and semi autonomous vehicles its full integration in our society is still deeply challenging Primarily there are crucial human elements at play while autonomous vehicles are not yet adopted in mainstream society efforts towards cyber physical vehicular systems should continue to identify and model factors that uniquely affect human in a vehicle An intelligent vehicle system must be aware of not only the surrounding environment and vehicles but also of how the environment might affect the agents i e human and non human drivers and inter agent interactions Safety and comfort of drivers have been improved over the decades through extensive studies in different driving conditions e g road terrains speeds etc but many exciting questions remain open For instance what are the effects of driver annoyance on both driving style and passenger comfort How do we identify and predict driver annoyance based on onboard sensor signals How then do we use these car signals to adapt the driving settings to increase the passengers This work was supported by NSF Award 154512 and by the Coordenac ao de Aperfeic oamento de Pessoal de N vel Superior Brasil CAPES Finance Code 88881 120236 2016 01 1E Araujo and E R Nascimento are with the Computer Science Department Universidade Federal de Minas Gerais Belo Horizonte Brazil edsonroteia erickson dcc ufmg br 2M Gregor is with the Department of Control and Information Systems University of Zilina Slovakiamichal gregor uniza sk 3I Huang and R Bajcsy are with the Department of Electrical Engineer ing and Computer Sciences University of California Berkeley CA 94270 USA isabella huang bajcsy berkeley edu comfort From being stuck in heavy traffi c dealing with other drivers that disobey rules to listening to the dissonance of honking cars annoyance on the road is almost inevitable Thus in this paper we evaluate the effect of driver annoyance in a real world driving study Our fi ndings are threefold First we identify the most predictive sensory car signals in discriminating between annoying and non annoying driving settings Second we investigate the effectiveness of such a classifi er and propose a method for real time estimation Third we measure the passenger s bodily dynamics using a pressure sensor mat to not only present a model of car passenger interactions but also to glean insight on passenger perceived comfort and smoothness of the ride II RELATED WORK Over the past several decades sound has been extensively studied as a natural and effective channel for inducing annoyance Auditory annoyance affects us psychologically by inciting selective attention inhibited memory lengthened reaction times and increased errors in cognition 1 2 It also leads to physiological effects such as increased blood pressure 3 hypertension 4 and other symptoms of stress 5 Moreover annoyance can beget undesirable physical reactions of frustration and anger 6 during driving tasks 7 which can lead to car accidents Most research in the auditory annoyance falls in quality evaluation engineering 8 or psychoacoustic metric de sign 9 This paper takes a step towards understanding how individual driver behaviors vary with auditory annoyance In general driving styles vary across personality traits 10 11 Paredes et al 12 presented a study of auditory induced stress and its effects in a driving simulator They showed that drivers arm muscles tighten when stressed and yield detectable differences in steering patterns In this paper we augment their work by performing a real world exper iment rather than in simulation building a more complete stress annoyance detection algorithm from an exhaustive set of in car sensors that is applicable not only during turns Furthermore we maintain that driving style is not only relevant as a function of driver state but also as an effector of passenger ride comfort In order to infer physical passenger comfort we must fi rst arrive at a model that captures the car rider interaction forces Experiments in assessing passenger comfort in public vehicles analyzed the effects of vibrations acceleration and jerk 13 14 but readings were taken from an inertial measurement unit attached to the vehicle rather than to the passenger We believe it is not only important to measure the vehicular forces during the ride but also to 2019 IEEE RSJ International Conference on Intelligent Robots and Systems IROS Macau China November 4 8 2019 978 1 7281 4003 2 19 31 00 2019 IEEE2234 Fseat Ffeet mg Fback may body x Fseat Ffeet mg mayFbelt body x Ffriction Ffrictio n z y b d max Fseat Ffeet Ffriction body y L mg z x a c Fig 1 Free body diagram in the transverse xz plane a at rest and b with labelled forces under lateral car accelerations Free body diagram in the sagittal plane yz plane of the passenger when the car is c accelerating and d decelerating in the y direction The green car seat indicates the placement of the pressure mat understand how exactly a passenger is affected by them To this end we utilize a Tekscan Body Pressure Measurement System mat to directly measure the interaction forces on the passenger s body as inspired by work in simulating driver forces for seat design in the work of Siefert et al 15 III DYNAMICS AND ACOUSTIC MODE CLASSIFIER MODELS We discuss our methods in modeling physical seat rider dynamics as well as an online classifi er for detecting different driving style modes from onboard vehicle sensors A Seat Rider Dynamics Model To best infer passenger comfort we need to understand the interaction forces between the car and the rider We apply the D Alembert method for describing inertial forces on the body induced by vehicle acceleration a to study the force torque relationships acting on the passenger separating the dynamics in the x y and z left forward and up directions We further assume that the only body joint relevant to the dynamics is the hip joint allowing only the rotation of the torso with respect to the fi xed lower body Thus all body torques are calculated with respect to this joint as the origin and the torso as the lever arm with the distance from the hip joint to the center of mass denoted as L For simplicity the body and seat are assumed to be rigid and the friction forces caused by the seat back is not modeled With m as the mass of the passenger and g as the gravitational acceleration we arrive at the following dynamical relationships Lateral Dynamics Lateral forces i e in the x direction induce side to side accelerations as depicted in Fig 1 a and b The dynamics equations that arise from the statics constraints P Fx 0 and P y 0 respectively are max Ffriction 1 body y maxLcos mgLsin 2 where Ffrictionis the static friction exerted by the seat and body yis the internal spinal muscle torque that opposes the counteracting moment The relationship between the lateral car acceleration a and the lean angle explicitly depends on body y Intuitively under lateral acceleration humans are neither perfectly limp body y remains at zero such that their torsos fl op over nor are they very stiff body yis extremely reactive such that they stay perfectly upright Rather humans remain relaxed until they have to react to a perturbation to balance 16 This reaction time which may differ for everyone is non zero Therefore for a given passenger larger accelerations yields a larger lean angle before the rider is able to properly stabilize themselves Since the pressure mat measures the forces applied on the seat by the human we expect that the shifting of the center of mass that accompanies any non zero lean angle will be observed in the pressure readings Forward Dynamics The dynamics for the yz plane are presented with reference to Fig 1 c and d Unlike in the transverse plane where lateral motion was symmetrical the sagittal dynamics differ between positive and negative y accelerations due to the asymmetry in the seat belt at the torso anterior and the seat back at the posterior In either case the static equations dictate that P Fy 0 and P x 0 Positive Acceleration y direction When the car ac celerates forward the passenger experiences an equal and opposite inertial force in the y direction shown in Fig 1 c According to the static equations may Fbacksin Ffriction 3 Fback 1 L body x maysin mg cos 4 where Fbackis the normal force exerted by the seat back Assuming that the passenger already makes full contact with the rigid seat back when at rest the car s accelerating will not change the pivot angle between the torso and the lower body Thus no observed changes in the pressure mat center of force is expected However the total force on the mat Fseatwill decrease when Fbackhas a non zero z component i e when the seat is not perfectly in the upright position Negative Acceleration y direction The passenger ex periences an inertial force in the y direction as depicted in Fig 1 d The statics equations dictate may Fbelt Ffriction 5 Fbeltcos 1 L body x mg sin maycos 6 where Fbeltis the backwards force applied on the passenger by the seat belt As with the lateral dynamics we expect that a larger acceleration magnitude ayyields a larger sagittal lean angle before the body is able to stabilize itself which subsequently results in a higher displacement of the pressure mat center of force in the y direction Vertical Dynamics The driver cannot directly generate forces in the z direction but they still arise via bumps and irregularities in the road Though not illustrated here a change in the total normal force recorded by the pressure mat Fseatwill arise when a non zero z acceleration azis present For example when the car acceleration is directed in the positive z direction we have the following relationship Fseat m g az Ffeet 7 where Ffeet is the force exerted on the feet by the fl oor 2235 B Online Classifi er for Acoustic Modes We propose a binary classifi er that given several seconds of driving samples distinguishes between two acoustic modes calm and annoying The motivation is to test how far the correlations that we have observed will generalize across dif ferent subjects We use the light Gradient Boosting Machine GBM classifi er 17 with 250 estimators The number of leaves the learning rate and the subsample ratio of columns were tuned using a bayesian optimization using the Tree Parzen Estimator 18 We feed the GBM classifi er with feature vectors consisting in 7 bin histograms with these six relevant measurements i linear acceleration x y ii angular velocity x z iii linear twist x iv brake torque request v throttle rate vi linear jerk x y Sliding Voting Mode For online classifi cation we must extract features not solely over entire driving trajectories but rather over smaller time windows of 2 000 samples i e 40s of data at the rate of 50Hz Since not every window will be informative enough of the acoustic mode we applied a sliding window the fi nal output at each point is determined by a vote among all the classifi er outputs that fall within that window Note that the windowing respects causality i e it is not centered but only considers past predictions IV EXPERIMENTS We now present our hypotheses and the details of the experiment designed to address them A Hypotheses H1 More annoying sounds leads to faster and jerkier driving We expect drivers to complete their drive more speedily to minimize time spent listening to annoying sounds In addition jerkier driving will arise from physical expression of annoyance induced anger 19 H2 Pressure mat readings support our seat rider dy namics model In particular we expect shifts in the center of force as recorded by the pressure mat to correspond with the x and y accelerations of the car Furthermore the total force registered by the mat should correspond with the z acceleration and positive y acceleration H3 Passengers comfort levels will be different under different acoustic modes Since acoustic modes impact driving style which directly impact the dynamical trajectory of the ride the riders should be able to physically detect these differences through the forces exerted on them throughout the ride These differences will be directly captured by the pres sure mat As a result passengers should feel less comfortable during the ride with annoying acoustic interference B Experimental Design We conducted a real world driving experiment at the UC Berkeley Richmond Field Station with 9 subjects with valid drivers licenses aged 21 to 31 6 male 3 female Each experimental trial involved two participants one driver and one passenger during which the driver performed three laps along the route in Fig 2 in a Lincoln MKS car equipped with a variety of onboard sensors During each lap one or none of Fig 2 Experimental driving route with coded segments of interest T1 to T8 are turning segments S1 and S2 are segments with stop signs and L1 and L2 are straight segments with no intermediate stops two soundtracks were played One of the soundtracks was a calming piece named Warm Darkness by Mia Strass1 and the other was of a crying baby2 These soundtracks were chosen to elicit divergent feelings of annoyance as justifi ed in 20 While the vehicle was in motion conversation was barred to avoid confounds Prior to the three laps drivers performed a test drive to familiarize themselves with the vehicle as well as properly memorize the route Manipulated variables The three laps per trial differed with respect to the auditory stimuli presented to the participants 1 No music was played in the vehicle 2 One of the soundtracks was played in the vehicle while the passenger listened to white noise with sound cancelling headphones 3 The other soundtrack was played in the vehicle while the passenger listened to white noise The ordering of the soundtracks was balanced across trials Dependent measures For each lap the car s sensory data were recorded in full In addition a 62 2 53 cm2Tekscan pressure mat with 1 sensel per cm2lined on the passenger seat measured the seat forces along the ride In addition after each drive that a soundtrack was played participants answered a series of open ended and Likert scale questions regarding their driving style and or perceived comfort Subject allocation Each subject participated in exactly two experimental trials once as a driver and once as a passenger Since the experimental roles are signifi cantly different we believe the second trials are not hindered by the subjects previous experiences Driver passenger pairs were assigned based on subject availability V RESULTS A H1 Auditory Annoyance Impacts Driving Style Objective measures The jerkiness of each lap was char acterized by a smoothness cost Sj Eq 8 equal to the integrated squared jerk normalized by duration inspired by the work in 21 Similarly we used a similar cost for acceleration Sa Eq 9 to quantify the speeds achieved along the route 1 2 2236 TABLE I P VALUES OF COST DIFFERENCES ACROSS DIFFERENT MODES CostAnnoying SilentAnnoying CalmSilent Calm Sj0 03310 01230 0198 Sa0 02580 03110 6702 Sj Rtf t ti x t 2 tf ti 8 Sa Rtf t ti x t 2 tf ti 9 a b Fig 3 The average cost metric across all laps a Sjand b Sa The average costs Sjand Safor all laps across the two different acoustic modes are compared in Fig 3 We included the silent mode as a baseline without acoustic interference We fi nd that the average costs are highest in the annoying mode and the statistical signifi cance of each pairwise comparison is presented in Table I Across all comparisons there are signifi cant differences at the 5 signifi cance level save for Sabetween the calm mode and the silent baseline In addition to comparing full trajectories we performed driving style comparisons with respect to specifi c segments of the route coded in Fig 2 Specifi cally we investigated how the speed acceleration and jerk profi les differed across the calm and annoying acoustic modes We deemed the distributions signifi cantly different if the p values for the Kolmogorov Smirnov and the Welch s t test were signifi cant at the 5 level For straight segments S and L type we found statistically signifi cant differences across acoustic modes only in the speed and y acceleration profi les with the mean of the absolute values of these measures being higher in the annoying mode Though insignifi cant differences in x acceleration were unsurprising due to the straightness of the path we also found that changes in neither the x nor y components of jerk were statistically signifi cant Virtually all turning segments T type yielded signifi cant differences in the x acceleration y acceleration and speed profi les The annoying mode yielded a higher mean speed a higher x accelerat