This paper illustrates the concept of “co-pilot” as an enabling technology for autonomous driving. A co-pilot system mixes the features of commercial Advanced Driver Assistance Systems (like blind spot, forward-collision warning, lane change assistant, overtaking assistant, and others) with human factors like driver distraction and intention. The copilot can provide a “suggested action” to the human driver through a dedicated Human-Machine Interface (a set of screens on the dashboard) or, alternatively, can be the enabling technology to build effective and user-friendly future intelligent transportation systems (i.e. Autonomous Driving Functions). We illustrate the results achieved by the European projects HoliDes and the next steps foreseen in the EU project AutoMate.
A number of intelligent agents is entering our lives and supporting us in a wider variety of tasks; in particular, this is definitely true for automotive domain, where automation in passenger cars is constantly increasing. In fact, current roadmaps of car-manufacturers and suppliers predict automated vehicles on highways by 2020. The reasons for that are threefold: • Zero Emission, with the reduction of fuel consumption and CO2 emission, as well as traffic flow optimization. • Demographic Change, including the support to unconfident drivers and the enhancement of mobility for elderly people. • Vision zero, which is the potential for more driver support by avoiding human driving errors (on which our paper is more focused).
Research in Intelligent Transportation Systems (ITS)
began in the late ‘80s with the PATH program in the US, the
PROMETHEUS project in the EU and the ASV projects in
Japan [Bishop, 2005]. Today, the development of highly
automated driving is the research focus of many OEMs and
research institutes, addressing the specific principles of
smart collaboration between humans and systems
Open questions regarding highly automated vehicles include the strengthening of driver’s sensing ability; the information in case of errors; and the reduction of the driving effort as well as an increased usability. Indeed human drivers are limited in recognizing, interpreting, understanding and operating in critical situations; moreover, they are prone to misbehaviors, drowsiness and distraction [HAVEit, 2013]. Nowadays, there are already on the market several
(e.g. blind spot, lane departure, emergency braking, semiautomatic parking, etc.) that are designed either to automate specific tasks or to provide additional information to the driver.
This research presents an artificial agent, named co-pilot, which provides a unique adaptive framework for supporting the human driver during critical situations or, alternatively, it can be regarded as an enabling technology for the Autonomous Driving Functions (ADFs). Indeed, the co-pilot is the core of such ADFs, by computing a (sub) optimal maneuver that takes into account both the lateral and longitudinal tasks under a common view. In addition the co-pilot is adaptive, namely the decision accounts for critical human factors, i.e. an estimate of the driver status (visual distraction) and intention. This is crucial to make the system response closer to human needs: for example, if the system detects that the driver is distracted, then it avoids to suggest more demanding maneuvers (such as a take-over).
The paper is organized as follows. Section 2 describes the technology we have considered and we are still using, inside the EU aforementioned projects. Section 3 describes the system architecture we adopted. Section 4 gives an overview of the current preliminary results. Finally, section 5 ends the paper, by providing the conclusions and illustrating the next steps of our research. 2
This section describes the methods and techniques that are used to realize the co-pilot, with a specific focus on the relevant human factors that enable the adaptation.
Driver intention recognition is mainly concerned with the recognition of maneuver (e.g. lane change) intentions, and a comparative review of works on maneuver intention estimation can be found in [Börger, 2013], [Doshi and Trivedi, 2011], [Kobiela, 2011], [Lefèvre et al., 2014]. The modeling landscape is mostly based on Dynamic Bayesian Networks (DBNs), Hidden Markov Models (HMMs) and their variants, or probabilistic and non-probabilistic discriminative models.
Intention recognition based on DBNs is usually organized
as follows: for each addressed maneuver, a distinct DBN is
learnt from a sequence of observations (training). The DBN
therefore models the dynamic evolution of the vehicle state
and/or position for the specific maneuver. Given a new
sequence of observations, the actual maneuver intention is
estimated by comparing the likelihood of observations for
each DBN
For intention recognition based on discriminative models,
commonly used techniques are Support Vector Machines
(SVMs), Multi-Layer Perceptrons or Logistic Regressions
Due to the sound foundation of machine-learning methods and the direct interpretability of their structure and parameters, we use DBNs for modeling driver intention recognition used in the co-pilot. In contrast to the aforementioned approach, we refrain from modeling the dynamic evolution of the vehicles state and position for different maneuvers in favor of a more direct representation of the statistical relationships between driver intentions and the complex traffic situation. We believe that this approach will provide for an earlier recognition of intentions solely based on the current environment (traffic, speed, car position) without the need for a lane change to have already started (i.e. without relying on the light indicators).
Driver distraction is a critical human factor with significant safety concerns [Regan et al., 2011]. Deriving knowledge on the human operator status can be very valuable for the operative system conditions. In this work we consider the following definition for driver’s distraction: “the diversion of attention toward a competing activity, which may result in insufficient or no attention at all to activities critical for safe driving”. Such a definition is quite general, and at the same time it allows us to capture the key elements of distraction, together with the important notion of insufficient or no attention being given to activities that are critical for safe driving. In practice, distraction can be split into visual and cognitive aspects. In this work we mainly consider the visual distraction, which is the diversion of attention toward a competing activity that requires the driver to look at a secondary target inside the vehicle instead of looking at the road.
In the literature several studies proved that visual
distraction can be successfully inferred using Machine Learning
(ML) approaches, that usually outperform other analytical
methods
A critical aspect needed to design adaptive autonomous systems is the decision making task, which has to weight several possibly conflicting data sources in order to decide a safe driving plan.
Errore. L'origine riferimento non è stata trovata.
shows the main building blocks of the car architecture,
where the co-pilot manages the automated functions
according to the situation and the driver needs, also taking into
account the environment constraints. The central point for
any automated systems is the ability to assess perception
and decision performance under a given condition in a
certain situation. With reference to the
perception-cognitiondecision process, as defined in [Stiller et al., 2007], input is
received from sensors (considering several aspects and
sources, e.g. internal camera for gestures and eye
movements, from maps, from the environment and so on) over
several processing steps via a geometrical-symbolic
representation of the current traffic environment to the generation
and control of suitable behavior. In this context, robustness
is essential: one successful method to obtain it is to consider
data-fusion from several sensors. This may happen on a
subsymbolic or symbolic level, in order to generate more robust
hypothesis. Thereby, it is crucial to not only propagate
knowledge through the cognition scheme but to augment
this knowledge with confidence measures, which are
consistently processed at each step of the cognition chain,
considering the confidence of previous processing steps along
with additional noise introduced by sensors and the
uncertainty introduced by the individual algorithms. Given that,
the co-pilot plans the safe maneuvers considering all these
factors and then distribute the shared maneuver execution to
driver and automation, including handing-over tasks to the
driver or accepting/rejecting tasks assigned by the driver to
the automation. In order to maintain the common frame of
reference
The co-pilot module is designated to support the maneuver decisions of the human driver, using a Bounded Markov Decision Process (BMDP) for the decision process. [Givan et al., 2000]. Figure 2 depicts the logical flow of the module. It starts by building the initial BMDP state s0 using the sensor data (world representation).
The set of actions Act considered in the prototype are: • Keep Your Lane (KYL): the EV (ego vehicle) continues following the current lane at the current speed. • Brake (Brk): the EV will try to brake (considering a span of possible decelerations). • Change Left Lane (CLL): the EV moves to the next lane on the left (considering a span of possible lateral accelerations) to start an overtake maneuver. • Change Right Lane (CRL): like before for the right lane, usually to conclude an overtake maneuver. • Slowdown (Slw): the EV decelerates following the current lane.
The projection function F↕︎(s, act) produces a new BMDP state s′ starting from state s and simulating the consequence of action act. This function follows the formulas of Errore. L'origine riferimento non è stata trovata., and involves vehicle dynamics and object kinetics. The projection function propagates uncertainty of the state parameters.
world representation driver distraction
(DDC) driver intention
(DIR) module input (1) trajectory
planning (2) trajectory rewards (4) derived intention (3) feasibility of strategies (5) solve
BMDP optimal strategy (6) strategy hysteresis
HMI output code (7) HMI realization The set of states is generated using a variation of an Online
Starting from s0, a tree of possible states is generated using F↕, evaluating all the actions up to time T. Each path in the tree is a trajectory. A sequence of actions (i.e. a policy) is mapped into a set of multiple trajectories, due to the uncertainty encoded by F↕. Step (2) assigns a reward to each trajectory, taking into account a safety measure of the state based on the standard time-to-collision measure:
R↕(s, act) = ramp(ttc, rmin, rmax) Where ramp(value, min, max) is a linear ramp function, and the reaction times used in the prototype are rmin=3, rmax=4 if the driver is not distracted, rmin=4, rmax=6 otherwise (hence distraction raises the human reaction time). In step (3) trajectories that exceed a safety threshold value are considered not safe. Step (4) considers the feasibility of the estimated driver intention, which could result in two outcomes: • Intended action is safe: this generates a (positive) suggestion in the HMI of doing that action, like: “you may change left”. • Intended action is not safe: this generates a (negative) indication in the HMI that the action is dangerous, like: “slowdown” because there is a vehicle ahead, or “do not change right” since the other lane is occupied.
The reward of a policy is the minimum of the rewards of that policy trajectories. The BMDP solution of step (5) consists in selecting the policy with the maximum reward function. The first action of the optimal policy is passed to the HMI module, after a hysteresis step (6).
In order to take an optimal decision, as illustrated in Figure 2, two blocks are also taken into account: the Driver
The DIR module is the software component designed to provide context-dependent estimates of the hidden lanechange maneuver intentions of the human driver. The hope is to be able to detect the intention prior to the turn indicators. This is an important aspect for the co-pilot, because – for example – it can learn the driver preferences during the “normal” driving and then adopt this during the automated driving task.
At runtime, the module receives input from the vehicle sensors (actuator states, current velocity, position data provided by a lane detection camera, surrounding obstacles as seen by the LIDARs). Within the DIR module, the available sensor information is first synchronized. An internal worldmodel based on particle filters [Koller and Friedmann, 2009] then augments the available information with better estimates of the environment and vehicle position and classifies surrounding vehicles according to predefined roles (e.g., the lead vehicle, the vehicle behind on the left lane, etc.). For actual intention recognition, the DIR module utilizes a DBN that describes the statistical relations among the intentions, the behaviors, and the information of the vehicles state and traffic situations: the DIR model. The details of the DIR module can be found in [Eilers et al., 2016] and in [Yan et al 2016].
The purpose of the DDC module is to classify the visual driver distraction based on vehicle dynamics data and internal camera, using the machine learning techniques described as following. This module provides information about the operator’s degree of distraction.
DDC consists of two components: the first component learns the classifier from a stream of sensory data. The second component uses the classifier to make prediction on the distraction status of the driver. The classifier can be trained either offline or online during its use. In this prototype, the classifier has been trained offline from system dynamics data collected from the prototype vehicle during 30 sessions. The 30 subjects drove for about 1 hour on normal and highway roads and they had to perform a SURT (surrogate task) while driving, in order to induce distraction. The vehicle dynamic variables considered in this study are: Speed [m/s] Lateral Position [m] Time To Lane Crossing [s] X,Y coordinates of front car (if any) Time To Collision [s] Lane Width [m] Position of accelerator pedal [%] Speed of car in front (if any) Heading Angle [deg] Road Curvature [%] Position of the brake pedal [%] Output of the monitoring system Steering Angle [deg] (head position and eyes tracking) Turn indicators [on/off] These values are directly available from the vehicle sensor data, or can be derived from those (e.g., time to collision is computed using the LIDAR data). The frequency of data collection is 20 Hz (1 data-point each 0.05s). Each of the continuous input variables above generates five input channels, namely the average, minimum, maximum, standard deviation and first derivative in a sliding window of given width. Discrete variables enter directly as input channels. The DDC module employs a SLFN network with 63 input neurons (one for each input channel), 100 hidden neurons and 2 output neurons. Weights are determined offline using ELM algorithm, and are loaded by the DDC module at runtime. The two output neurons generate the distracted/nondistracted probability distribution, which is then discretized to obtain the distraction classification used as input for the decision module.
Figure 3 shows how the system works in practice on the prototype vehicle with two small examples. These examples are extracted from a test drive done with the prototype vehicle on a highway near Torino, Italy. The purpose is to illustrate the adaptation on the driver intention/distraction in the decision process of the whole system. In example A the driver is fast approaching a slower car on the right lane. The DIR module infers that the most probable intention of the driver is to overtake that car, and the DDC considers the driver to be attentive. With this setting, the co-pilot module verifies that the overtake is safe (no obstacles), and suggests the driver with 7 seconds in advance that (s)he may overtake, supporting all the maneuvers. The suggestion does not depend on the turn indicators of the car. At stage A.4 in Figure 3 the DIR module infers that the driver should return to the right lane. The co-pilot module first shows a “keep your lane” enforcement signal until the right lane is occupied (A.4), and finally, shows the CRL message to support the reentrance. Note that a forward-collision-warning and a blindspot will behave differently. They would both signal the longitudinal and lateral dangers (FCW at A.3 and blindspot at A.5), because they are not adaptive to the overtake intention.
Example B shows a similar scenario where the driver is again fast approaching a slower truck, but the left lane is occupied by another car, and the DDC module considers the driver to be inattentive. In this case the co-pilot module determines that the safest (highest reward) policy is a slowdown action, and it emits a “slowdown/do not change left” warning to the driver. If the driver does not respond quickly, an emergency brake signal would appear. After having adjusted the speed to follow the truck, the DDC module determines that the driver is attentive again (unclear whether attention is deduced because the speed has been reduced or because the module says so and in this scenario we consider the case in which the driver is attentive again at this point in time, and a CRL message is shown to suggest overtaking. 4
A prototype has been developed to test the feasibility of the co-pilot. The prototype addresses the driving in a highway scenario, which is adequate since the focus of the project is the adaptation to the human behavior more than the adaptation to the environment. The prototype has been realized in two forms: a vehicle and a simulator. Both runs the same software stack built on the RT-Maps 4 framework1.
In order to test the feasibility of the system, including the co-pilot, DIR and DDC modules, we have considered one prototype (as illustrated in Figure 4), which is adequate for the adaptation to the human behavior and to the environment. The prototype has been realized in two forms: a vehi[1] 1 RT-Maps framework. https://intempora.com/. cle and a simulator. Figure 4(left) shows the sensors used for the data collection: 1) an external camera used by the lane-detection algorithm (to build the road model); 2) an internal camera used to scan the human driver face; 3) four LIDAR sensors that detect environment obstacles. Figure 4(right) shows the setup of the simulator environment, which includes 4) the distributed co-pilot HMI; 5) the SURT used to distract the users. The simulator runs the SCANeR II software.
A large number of experiments was carried out to test the DDC module, by varying the learning algorithm parameters, such as the number of neurons, learning rates, number of training instances, etc. Moreover, the collected data have been averaged over a period of time that varies, in the various experiments, between 1s and 2s. In order to be consistent with the target variable (distracted or not-distracted), data have been labelled distracted when the driver was not looking at the road for the whole considered period (using the internal camera of the vehicle), not distracted otherwise. Table 1 shows the main results:
98.99+/-0.04
87.52+/-1.37
0.07+/-0.02 0.18+/-0.02
The table reports the average classification rates (accuracy, false negative rate and false positive rate) obtained by training the DDC module using a leave-one-out strategy. In detail, we selected 25 drivers out of the 30 ones for which we collected data (5 were discarded due to a small number of distraction cases). In turn, one of these dataset has been used for testing while the other 24 have been combined together to form the learning set, and the process repeated for every dataset (25 runs in total).
In details, the first two columns report the average accuracy on the training sets and test set, respectively. The accuracy on the test cases shows good generalization ability towards new drivers. For what concerns the false negative rate, it should be noted that in practical situations there are many more cases of driver not distracted than there are cases of distraction, and the latter are more difficult to identify, in general, because good drivers tend to drive safe even when partially distracted.
The evaluation of the DIR module has been the subject of a separate analysis. Interested readers can find it in [Eilers et al., 2016] and in [Yan et al 2016].
The aim of the system experimental assessment was to objectively measure its performance against a state-of-the-art baseline system. The assessment was done on the driving simulator, for safety reasons. The baseline system includes both a blindspot and FCW systems for lateral and longitudinal warnings.
Model@143.639D 189.025 The test involved 30 subjects (15 men, 15 women; average could be cMaoudsele@d158b.y554tDhe f7a9.c4t34t2hat the scenarios make these age: 39 years; average years of driving license: 20; average braking necMeodsesla@r1y61.d4u41eD to t49h91..e7996c56r95itical situations that happens. km/year: 15.660) on five tracks, without anticipating the PI4 takes inMotdoela@c21c7o.u36n4tD the effect of the recognition of driver kind of support system that they would have seen while In[44]:= iGnrtied@n8t8iBoonxW:hwiskheernChatrhte@PDI2IDR,BomxWohdisukleerCihnarfte@rPsI4tDh,eBoixnWtheinskteiorCnhaortf@PoIv5De<r<-D sdirmivuinlagt.orA),fttehre atewstairsmp-eurpfotrrmacekd (ttwoog etitmceosnfoindetnwcoe twraicthksth(5e ltGGaarrnkiiddBie@@no88xcg88WhDB,hioiasxtsntWhkrhgeeiirebsCukhstemayirrsoCtanht@neCaPhreImat4u@r,Ptv"s@IeOP2uurI,gt2"ilDgOni,ueetDrsalistdis"setDvrr,atsiBh"bnoDuecx,tWiehominssCaekhneearrmeCtuh@saPvrIett4o@rDP;,Im5s,ou"dgOigufteylsitetihrnse"gDa<t<thtDieminutes each), with the baseline and with the prototype sys- tudDeisotfritbhuetiounsCeharrst@inPI5cDh<<aDnging the lane in advance, increasing tem. Half of the participants see the system after the base- tGhreid@s8a8fHeitsytodgriasmt@aPnI1cDe,.Histogram@PI3D<<D line, while the other half see it before (to randomize and PI2 PI4 PI5 avoid bias).
The two test tracks are two-lane highway scenarios designed Out[44]= to create specific dangerous situations to the driver, like: a preceding vehicle that brakes unexpectedly, a slow preceding vehicle with an incoming car on the left lane, … Details bbaasseelliinnee AsdysCteomS bbaasseelliinnee AsydsCteomS bbaasseelliinnee AsydsCteomS on the tested situations can be found in [HoliDes, 2016] Figure 5: Box-plots of the continuous PIs. (D9.9). Drivers were requested to perform a SURT task that activated at random every 30/45 seconds on a secondary Out[45]= Figure 5 shows the box-plots of the three continuous PIs screen that is located on the side of the simulator screen (to (left is baseline, right is new system). The statistical test trigger a visual distraction). shows that all PIs have enough samples to assess for a staThe following Performance Indicators where considered: tistical difference, except for PI4 (even if it is showing PI1. Number of accidents occurred during the test. slightly different distributions), which requires more samPI2. Percentage of driving time where the TTC of the pre- ples.
ceding vehicle is less than 2 seconds. Out[46]= PI3. Number of times the driver presses the brake strongly, 5 Discussion and Conclusions achieving a sudden hard braking (deceleration of more In the European co-funded project HoliDes [HoliDes, than -8m/s2). 2016], a first version of the co-pilot has been implemented PI4. Average distance to the preceding vehicle when the and used in a limited form to produce a comprehensive preuser performs the lane change. The purpose is to ventive safety system capable of giving information and measure if the new system increases safety, avoiding Out[47]= warnings only. However, this concept of the co-pilots are the Peltzman effect [Peltzman, 1975] (inducing confi- potentially suited to more sophisticated applications, such as dence in making risky maneuvers). the ones developed in the EU project AutoMate (http://www.automate-project.eu/). In fact, if necessary, the PI5. Average TTC when the driver starts pressing the brake, driver could “loosen” control and let the system autonoto measure the impact of slowdown/brake suggestions. mously navigate, or can “tighten” control and reclaim auTable II shows the average PI values obtained from the user thority. On the other way around, if necessary, the system tests, and the statistical significance. The values show fa- may be programmed to completely take over from the driver vorable performance of the new system over the baseline. in certain conditions.
PI1 PI2 PI3 PI4 PI5 In all cases, the important aspect is that the co-pilot will Baseline 0.1724 0.0126 1.3103 73.5552 2.7134 be adaptive and cooperative, thus the driven vehicle should co-pilot 0.0862 0.0069 1.3793 82.1026 3.3742 appear to be driven by a human, being easily interpretable to Signif. <0.0001 <0.0001 0.0011 0.5461 0.0085 other human road users. In this context, the co-pilot is the Table II: PI results of the evaluation. enabling technology for implementing ADFs, with the capacity to improve the human–vehicle interactions, by conFor the number of accidents, they are strongly reduced: in sidering and exchanging intentions between co-driverthe baseline, there are 10 accidents in the test cases, while in equipped vehicles, as well as by taking into account the the new system there are only 5 accidents. This means the driver state (i.e. visual distraction). co-pilot was able to halve the number. A similar improve- In this paper, we have attempted to set out a viable ment has been achieved also for PI2, where the time spent roadmap for producing the co-pilot enabling technology, by the drivers in potentially critical situations has been re- making significant use of recent developments in cognitive duced by around 50%. systems, in order to address the adaptation and continuous The higher value of the system on the indicator PI5 means support to the human driver. This system is realized using a that, when the driver starts braking, the TTC is greater in the decision process that balances multiple action outcomes new system than it is in the baseline, meaning that the co- with the inferred human status (intention, distraction). This pilot increases the awareness of the driver to dangerous situ- ipcraoldmuceesssacgoenstethxrtouualgizhead dsetrdaitceagtieeds HthMatI aorne tshheowvenhiacslegdraapshh-ations. The number of hard brakings is almost equal, which board, or alternatively, can be regarded as a “virtual driver”, able to take the vehicle control and thus implementing the ADFs.
Moreover, other achievements will need further research at the intersection between cognitive sciences and intelligent vehicles. In particular, we plan to investigate the impact of introducing a human-like behavior of the co-pilot, sharing experience and roles with human drivers. In this context, the system could use its emulation capacity (in a way similar to human rehearsing of possible experiences) to discover and learn higher-level behaviors that might prove more effective. This enables the co-pilot to become an expert driver without directly needing training examples from expert humans. In this sense, reinforcement-learning techniques would probably also allow to better tailor such a system to each human driver.
This work was supported by the EU Artemis Joint Undertaking research project HoliDes, grant no. 332933. [HoliDes, 2016] project: deliverables
http://www.holides.eu/public[Börger, 2013] J. Börger, “Fahrerintentionserkennung und Kursprädiktion mit erweiterten Maschinellen Lernverfahren“, dissertation, Universität Ulm, 2013. [Doshi and Trivedi, 2011] A. Doshi and M. M. Trivedi, “Tactical Driver Behavior Prediction and Intent Inference: A Review”, in Proceedings of the 14th International IEEE Conference on Intelligent Transportation Systems, 2011, pp. 1892-1897.