assessment and design of the Stop-and-Go ADAS ... current Stop-and-Go control algorithms by undertaking .... System can't cope â very strong lead vehicle.
ASSESSMENT OF THE STOP AND GO FUNCTION USING REAL DRIVING BEHAVIOUR Greg Marsden, Mark Brackstone, Mike McDonald Transportation Research Group, University of Southampton, Highfield, Southampton, U.K. Proc. of the ADAS 2001 Conference, Birmingham, U.K.. IEE Conference Publication 483, IEE, London, Sept. 2001. pp 76-80. Copyright IEE, London, U.K.
Although clearly the next step in ADAS design evolution, and already the focus of several prototypes, remarkably little evaluation (3, 4, 5) has been undertaken of the features that Stop&Go should possess, or how beneficial it could prove.
ABSTRACT This paper reports on initial findings concerning the assessment and design of the Stop-and-Go ADAS (Advanced Driver Assistance Systems) functionality, which allows low speed distance keeping between successive vehicles. The work is part of the Motorway Operations project being undertaken in the UK, dealing with understanding driver behaviour, and assessing and designing existing and new types of ADAS.
To this end an evaluation has been undertaken of two variants of a Stop&Go control algorithm through their comparison with real driving speed profiles collected in heavy flow conditions. In the next section, the methodology for the evaluation and the control algorithms are introduced. Next, a brief summary of the empirical data used is provided before the results of the comparison are presented. The final sections draws some preliminary conclusions from the findings.
The research presented examines the suitability of current Stop-and-Go control algorithms by undertaking a comparison between output from a micro-simulation model and time series data collected using an instrumented vehicle deployed during rush hour on the A35 in Southampton in the U.K. The paper focuses on the impact of differing acceleration/deceleration policies, as well as examining how the algorithm copes with the sharp decelerations required during shockwaves. The paper concludes that unassisted drivers react earlier than the current algorithm, perhaps due to anticipation. More complex algorithms may need to be applied to match human performance in the complex low speed environment.
METHODOLOGY The overall evaluation follows one of the methodologies introduced in earlier evaluations of ACC (1). This method uses a 2-car microscopic model (6), where the lead vehicle is driven according to a set speed profile obtained from empirical data. The motion of the following vehicle meanwhile is controlled by a set algorithm, while initial conditions, eg following gap time headway, relative speed and absolute speed are set to match the initial values of the following vehicle from the empirical data. A comparison may then be undertaken of how the human driver data compares with that which a control algorithm would produce.
BACKGROUND The (partial) automation of vehicle longitudinal control based on sensing the relative motion of the preceding vehicle has recently come to fruition with the release by several manufacturers of ACC (Adaptive Cruise Control) equipped vehicle. The functionality of ACC is based on the desire to keep a set time gap between vehicles, and as such is characterised by a range of presets such as a target headway (in seconds), a maximum level of decelerative authority (the maximum level at which the system can decelerate the vehicle before human intervention is required), and a minimum operable speed (kph), below which the ACC cuts out and human control must be resumed.
The control algorithm is taken from the literature (4, 5, 7). It is based on a simple linear controller law with: a = k1*DV + k2*(DX – DXdes) with ‘a’ the acceleration or deceleration of the S&G equipped vehicle, DV the relative speed between it and the preceding vehicle, DX the space gap, and DXdes, the desired spacing given by Th*v, a set gap time headway multiplied by the vehicles speed, ‘v’. k1 and k2 are constants.
A number of studies have examined the effects of differing combinations of these three variables on motorway traffic (1, 2). Although many questions remain unanswered related to ACC usage, attention is now turning to the design of a system with no (or a very low) minimum speed cut-out that would see use by vehicles in queues, the ‘Stop&Go’ function.
EMPIRICAL DATA Data used as a basis for simulation comparisons were collected in Nov. 2000 in the evening peak on the A35 westbound heading out of Southampton by an instrumented vehicle. The vehicle used (8) is equipped
1
with an automotive radar and is able to measure the distance between the equipped vehicle and (in this case) a following vehicle. 81 time series were collected over seven peak periods with an average length of 127 sec. (an example of which is given in Figure 1). For the purposes of these investigations 9 files of longer than 3 minutes duration were selected for use. All traces contained periods of zero speed. The characteristics of these traces are summarised in Table 1. Ave. max. V (kph) 73.22
Combined ave. V (kph) 29.35
Ave. max. 2 decel. (m/s )
Ave. max. 2 accel. (m/s )
-1.92
1.89
4.
observed driver. Under some conditions, the Stop&Go system can not respond quickly enough. Different control strategy – the Stop&Go controller tries to work to a time headway control of 1.4 seconds (where DXDes is calculated as 1.4*v). In reality, the driver is able to select a following strategy dependent on the conditions which may be different. Differences between the observed following strategy and the Stop&Go controller sometimes lead to interventions.
Table 2 below summarises the number of each of the intervention events. Intervention type N° of interventions
Table 1: Summary of time series traces RESULTS
1 7
2 14
Table 2: Characteristics of interventions – Parameter Set 1
The results section presents the output of the comparison of the algorithm performance against the observed human performance for two different algorithm parameter sets. In Parameter set one, k1 = 0.3 and k2 = 0.03, whilst in Parameter set 2 k1 = 0.6 and k2 = 0.03. In both cases Th is set to 1.4 seconds. The main outputs of interest are: number of modeled driver-system • The interventions; • Acceleration/deceleration performance; • Relative velocity; and • Time gap headway.
3 6
4 3
driver-system
The controller algorithm uses two terms to calculate the change in speed. The first is a straight function of relative velocity. When the following vehicle is closing (negative relative velocity) this term request deceleration. However, evidence from the intervention events shows that the observed driver begins harder decelerations at relative velocities closer to zero than the algorithm provides (Figure 2). This may be a function of anticipation based on surrounding traffic conditions or due to driver anxiety although neither of these statements can be proven. However, this lack of response results in over 50% of non type 1 interventions.
A description of driver-system interventions is provided separately for the two parameter sets before a joint comparison of modeled against real following performance is presented.
Type 4 interventions occur when the Stop&Go algorithm is controlling at a different strategy to the manual follower. When the Stop&Go algorithm is following more closely, occasions arise where it is not able to cope with a lead vehicle deceleration. Whilst fixed time headway has been employed as a basis for control in Adaptive Cruise Control systems (2) it does not seem to reflect driver practice at low speeds. This is discussed further in the combined results analysis.
Parameter Set 1 Parameter set 1 exhibited a large number of modeled driver-system interaction, reflecting points where the driver model within the microscopic simulation felt that the vehicle was outside a comfortable operating zone for the driver whilst under Stop&Go control. Of the 9 following traces, 6 were found to contain a total of 30 interventions. A closer examination was made of each intervention event and the events classified as one of four types: 1. Model-system problem – typically occurred at very low speed where the vehicle was about to stop. These are related to the calibration of the simulation model as the vehicle approaches to a stop rather than a dangerous traffic situation. 2. Lag of acceleration/deceleration – the Stop&Go system responds more slowly than the driver to lead vehicle deceleration and more negative relative velocities build up leading to driver intervention. 3. System can’t cope – very strong lead vehicle decelerations which provoke strong reactions in the
The second term in the control algorithm modifies the speed to try and achieve following distance equal to the desired following distance (DXDes). However, at low speeds, DX-DXDes is rarely more than 10m and the maximum that this term could contribute to deceleration 2 is therefore around –0.3m/s . At the beginning of lead vehicle decelerations, this term can act in the opposite direction to the first term, contributing further to deceleration lag. Parameter Set 2 Parameter set 2 doubles the accelerative influence of relative velocity compared to Parameter set 1. The same traces were run through to determine whether a stronger
2
relative velocity influence could reduce the number of interventions across all categories. Table 3 shows the intervention results. Intervention type 1 2 N° of interventions 0 0 Table 3: Characteristics of interventions – Parameter Set 2
traces for Parameter set 1 (p