Probabilistic Prediction from Planning Perspective: Problem Formulation, Representation Simplification and Evaluation Metric

Abstract

Accurate probabilistic prediction for intention and motion of road users is a key prerequisite to achieve safe and high-quality decision-making and motion planning for autonomous driving. Typically, the performance of probabilistic predictions was only evaluated by learning metrics for approximation to the motion distribution in the dataset. However, as a module supporting decision and planning, probabilistic prediction should also be evaluated from decision and planning perspective. Moreover, the evaluation of probabilistic prediction highly relies on the problem formulation variation and motion representation simplification, which lacks a summary in a comprehensive framework. To address such concerns, we provide a systematic and unified framework for the analysis of three under-explored aspects of probabilistic prediction: problem formulation, representation simplification and evaluation metric. More importantly, we address the omitted but crucial problems in the three aspects from decision and planning perspective. In addition to a review of learning metrics, metrics to be considered from planning perspective are highlighted, such as planning consequence of inaccurate and erroneous prediction, as well as violations of predicted motions to planning constraints. We address practical formulation variations of prediction problems, such as decision-maker view and blind view for viewpoint, as well as reactive prediction for interaction, so that decision and planning can be facilitated.

Non-local Social Pooling for Vehicle Trajectory Prediction

“…Numerous vehicle motion prediction models have recently been proposed. Lefèvre et al [17] and Zhan et al [18] consider that one of the main distinctions between vehicle behavior forecasting methods is whether or not they take into account interactions between the surrounding vehicles.…”

Towards a Fatality-Aware Benchmark of Probabilistic Reaction Prediction in Highly Interactive Driving Scenarios

“…However, it often denotes either the desire in the human mind or the executed motion pattern, which is not explicit. As was pointed out in [13], it is impossible to obtain the strict ground truth of the desire in the human mind, especially in highly interactive driving scenarios. We will use the following example of ramp merging on I-80 in NGSIM dataset [19] to explain.…”

“…In [13], two spatiotemporal representations were summarized, namely, prototype trajectory and reachable set, which can incorporate the generated motion patterns. In this paper, we adopt prototype trajectory as the spatiotemporal representation.…”

“…At the inference stage, given a sequence of historical motion, we can obtain the likelihood of the observation sequences for each HMM by the forward algorithm. Then the likelihood values are normalized to obtain the posterior probability of each situation, which can be written as p k ′ , where k ′ = 1, ..., K, and K is the number of possible situations according to the combination of pass-yield motion patterns [13]. For each GMM, we can obtain the probabilistic density for the jth prototype trajectory, denoted as f k j .…”

“…In [13], the under-explored aspects of probabilistic prediction were highlighted with a comprehensive review on problem formulation, representation simplification and evaluation metric. In this paper, we provide a preliminary solution for the problems pointed out in [13], and implement the ideas with evaluation results as a mini benchmark. The main contribution of this paper is to propose a fatalityaware evaluation metric for probabilistic reaction prediction in extremely challenging driving scenarios with interaction.…”