RiskMap: A Unified Driving Context Representation for Autonomous Motion Planning in Urban Driving Environment

TL;DR

The results illustrate that the RiskMap generation tools and model structures can improve driving safety and smoothness, and the limitation of the method is also discussed.

Abstract

Motion planning is a complicated task that requires the combination of perception, map information integration and prediction, particularly when driving in heavy traffic. Developing an extensible and efficient representation that visualizes sensor noise and provides basis to real-time planning tasks is desirable. We aim to develop an interpretable map representation, which offers prior of driving cost in planning tasks. In this way, we can simplify the planning process for dealing with complex driving scenarios and visualize sensor noise. Specifically, we propose a unified context representation empowered by deep neural networks. The unified representation is a differentiable risk field, which is an analytical representation of statistical cognition regarding traffic participants for downstream planning tasks. This representation method is nominated as RiskMap. A sampling-based planner is adopted to train and compare RiskMap generation methods. In this paper, the RiskMap generation tools and model structures are explored, the results illustrate that our method can improve driving safety and smoothness, and the limitation of our method is also discussed.

Figures (3)

• Figure 1: Planning in urban scenarios is a challenging task because it needs to consider the perception, prediction, and control uncertainties together with complex road environments. Our method encodes and represents all perception information in a risk map, which is a middleware of sensing and planning. The purple rectangle represents the ego vehicle, the red blocks are other agents. The transparency of orange-yellow fields represent the occupancy probability of other traffic participants. The white-yellow-orange points represent the risk value of sampled positions. Sample points with deeper shades of red signifying more severe risks.
• Figure 2: Illustration of the proposed framework of map generation and planning module with their training process. The framework is roughly divided into two modules: prediction and planning. Prediction: Reference lines, map elements, ego & agents motion, and geometry are encoded with MLP/GRU modules. Global attention is used to model the interaction between objects. The multi-modal MLP decoder generates a series of multivariate normal distribution parameters to represent the vehicle future positions. Planning: A sampling-based planner is adopted, and all the sampled spatial-temporal points are with distance information to representative represent driving contexts. The global encoded latent variable in the prediction module generates parameters mapping distance to risk value space. Risk values are concatenated with collision risk. The trajectory with the minimum risk is selected as the planning result. Training: A two-stage supervised learning process is adopted. The predictor and planner are trained with listed metrics, respectively.
• Figure 3: Demonstration of SeqMVN, deeper color represents later time and the transparency represents the occupancy probability. The image shows that the multi-modality of SeqMVN helps to fit the complex shape of the probabilistic occupancy map. And the Gaussian distribution at different time points in each modal is correlated.