Behavior prediction at multiple time-scales in inner-city scenarios

Abstract

Abstract-We present a flexible and scalable architecture that can learn to predict the future behavior of a vehicle in inner-city traffic. While behavior prediction studies have mainly been focusing on lane change events on highways, we apply our approach to a simple inner-city scenario: approaching a traffic light. Our system employs dynamic information about the current ego-vehicle state as well as static information about the scene, in this case position and state of nearby traffic lights.Our approach differs from previous work in several aspects. First of all, we hold that predicting the precise sequence of physical and actuator states of a car driving in dynamic innercity traffic is both challenging and unnecessary. We therefore represent predicted behavior as a sequence of few elementary states, termed behavior primitives. As a second aspect, behavior prediction is treated as a multi-class learning problem since there are multiple behavior primitives. Rather than causing problems, we show that this fact can be exploited for computing information-theoretic measures of prediction confidence, thereby allowing to identify and reject unreliable predictions.We show that the horizon of predictions can be extended up to 6s, and that uncertain predictions can be detected and eliminated efficiently. We consider this a significant result since typical prediction horizons are usually in the range of 1 to 2s.The main message of this paper is that simple learning methods can achieve excellent prediction quality at long time horizons by operating purely on the "system-level", i.e., using an abstract, low-dimensional situation representation. Since the learning approach greatly reduces the design effort, and since we show that the prediction of multiple behavior classes is feasible, we expect our architecture to be scalable to more complex scenarios in inner-city traffic.

A survey on motion prediction and risk assessment for intelligent vehicles

“…Here, a maneuver is defined as "a physical movement or series of moves requiring skill and care" [23]. The word "behavior" is sometimes used in the literature for the same purpose [24][25][26][27][28], but for the sake of clarity the word "maneuver" will be used throughout this paper. Trajectory prediction with Maneuver-based motion models is based on the early recognition of the maneuvers that drivers intend to perform.…”

“…Context and heuristics can be used to determine what maneuvers are likely to be performed in the near future in a deterministic manner [39]. For classifying maneuvers in more complex scenarios, discriminative learning algorithms are very popular, such as Multi-Layer Perceptrons (MLP) [28] Logistic regression [40], Relevance Vector Machines (RVM) [41], or Support Vector Machines (SVM) [42][43][44]. An equally popular alternative is to break down each maneuver into a chain of consecutive events and to represent this sequence of events using a Hidden Markov Model (HMM).…”

Surrogate Safety Measures Prediction at Multiple Timescales in V2P Conflicts Based on Gated Recurrent Unit

“…Compared to the existing literature that has mainly focused on predicting the trajectories (or intentions) of the ego-vehicle and/or other surrounding road users [9], the developed approach differs not only by modeling safety parameters directly related to the V2P interaction severity levels but also by the fact that the multi-step-ahead prediction depends on a low-dimensional representation of the current situation: the calibrated system depends only on six parameters related to driver mobility features and traffic scene properties. Indeed, Ortiz et al [13] proved that there is no need to employ the state of the actuators as features for predicting future behavior: the authors have actually predicted with simple learning algorithms (i.e., multi-layer perceptron neural networks) the braking behavior of drivers approaching a traffic light with very good accuracy at time scales up to 3 s, using as input features only the ego-vehicle speed, state, and distance to the nearby traffic light. In addition, the approach presented in this paper has some aspects in common with the research of Zhang et al [23], but the authors predict the instantaneous severity level (based on TTC and PET) of vehicle-pedestrian interactions whose dynamics are captured by fixed cameras placed at signalized intersections, whereas the aim of the current study is a multi-step-ahead prediction on a running vehicle.…”

“…Existing techniques for learning driver or other road user behavior sequences from the set of features acquired by the vehicle sensor system can be divided between two methods: classification and regression. Classification problems concern the identification of movement intention labels, which are also called "behavior primitives" [13]: these classes segment complex driving behavior into a sequence of basic elements, such as lane keeping, left/right lane change, left/right turn, go straight, or speed maintenance, braking, and stopping. In the latter context, Khairdoost et al [14] implemented a deep learning system that can anticipate (by 3.6 s on average) driver maneuvers (left/right lane change, left/right turn, and go straight), exploiting the driver's gaze and head position as well as vehicle dynamics data.…”

Safety of Lane Change Maneuver Through A Priori Prediction of Trajectory Using Neural Networks

“…However, the accuracy and validity depends mainly on the drivers' behavior during lane change. Majority of the accidents on highways take place during lane change maneuvers which obstruct the traffic flood and create an overload in addition to loss of precious human life [4], [5]. The advances in technology have made it possible to simulate the complex models of highway traffic and give the warning for safe driving on the highways during the lane changing and overtaking.…”