Improving Autonomous Driving Safety with POP: A Framework for Accurate Partially Observed Trajectory Predictions

TL;DR

Partial observations in urban trajectory forecasting create safety risks. The paper introduces Partial Observations Prediction (POP), a three-stage framework that first learns robust history representations via self-supervised learning and then transfers that knowledge to a student model through feature distillation from a fully-observed teacher. POP demonstrates competitive open-loop accuracy against state-of-the-art methods and significantly improves safety in closed-loop, interactive simulations, reducing collisions and promoting safer interactions with surrounding traffic. This approach offers practical robustness to sensing and tracking gaps, improving reliability of autonomous driving systems in real-world, partially observed environments.

Abstract

Accurate trajectory prediction is crucial for safe and efficient autonomous driving, but handling partial observations presents significant challenges. To address this, we propose a novel trajectory prediction framework called Partial Observations Prediction (POP) for congested urban road scenarios. The framework consists of two key stages: self-supervised learning (SSL) and feature distillation. POP first employs SLL to help the model learn to reconstruct history representations, and then utilizes feature distillation as the fine-tuning task to transfer knowledge from the teacher model, which has been pre-trained with complete observations, to the student model, which has only few observations. POP achieves comparable results to top-performing methods in open-loop experiments and outperforms the baseline method in closed-loop simulations, including safety metrics. Qualitative results illustrate the superiority of POP in providing reasonable and safe trajectory predictions.

Figures (5)

• Figure 1: Partially observed predictions in real-world situations. In this scenario, the self-driving car is making a left turn, but another car is accidentally turning right from the left turn lane. Due to insufficient observations, the future trajectories provided by the prediction algorithm fail to include this possibility, leading to a dangerous situation.
• Figure 2: Overview of POP. Our method consists of three stages. The motion forecasting stage involves training a teacher model with complete observations. The SSL stage consists a mask procedure and a history reconstruction pre-task. During the distillation stage, the teacher model's parameters are frozen, and a feature distillation strategy is applied to the hidden features.
• Figure 3: Distribution of observations with standard 20 frames. The gray bars represent the detection of each frame for a fixed observation period, while the blue bars represent tracking. The Red indicates tracking failures.
• Figure 4: Observation length evaluation analysis on Av1 and Av2. Methods that are evaluated on Av2 are marked with symbol "*".
• Figure 5: A collision scenario caused by limited observations. The AV (red) is surrounded by yellow vehicles, the blue scatter line is the planned trajectory of the AV, and the predictions of the AV for other vehicles are marked with scatter lines of other colors. Due to limited observations, the HiVT predictor inaccurately predicted the future trajectory of a vehicle (black circle), causing the AV to accelerate and ultimately collide. In contrast, the AV equipped with the POP-H predictor exhibited superior predictions, ensuring safety.