[Paper Review] Motion Estimation in Occupancy Grid Maps in Stationary Settings Using Recurrent Neural Networks

Marcel Schreiber, Vasileios Belagiannis+2 more|arXiv (Cornell University)|Sep 25, 2019

Autonomous Vehicle Technology and Safety28 references21 citations

TL;DR

This paper proposes a recurrent neural network (RNN) with ConvLSTM layers to estimate dynamic occupancy grid maps from sequences of lidar-based measurement grid maps, predicting both occupancy probability and 2D velocity per cell. The method outperforms particle filter baselines in complex maneuvers like braking and turning, delivering more consistent and accurate velocity estimates with fewer erroneous predictions in static areas.

ABSTRACT

In this work, we tackle the problem of modeling the vehicle environment as dynamic occupancy grid map in complex urban scenarios using recurrent neural networks. Dynamic occupancy grid maps represent the scene in a bird's eye view, where each grid cell contains the occupancy probability and the two dimensional velocity. As input data, our approach relies on measurement grid maps, which contain occupancy probabilities, generated with lidar measurements. Given this configuration, we propose a recurrent neural network architecture to predict a dynamic occupancy grid map, i.e. filtered occupancy and velocity of each cell, by using a sequence of measurement grid maps. Our network architecture contains convolutional long-short term memories in order to sequentially process the input, makes use of spatial context, and captures motion. In the evaluation, we quantify improvements in estimating the velocity of braking and turning vehicles compared to the state-of-the-art. Additionally, we demonstrate that our approach provides more consistent velocity estimates for dynamic objects, as well as, less erroneous velocity estimates in static area.

Motivation & Objective

• To address the challenge of modeling dynamic vehicle environments in complex urban scenarios using occupancy grid maps.
• To improve velocity estimation accuracy for moving objects, especially during braking and turning maneuvers.
• To reduce erroneous velocity predictions in static regions compared to particle filter-based methods.
• To develop a learning-based approach that does not rely on motion assumptions like constant velocity.
• To enable robust, consistent motion estimation using spatio-temporal sequences of lidar measurements.

Proposed method

• The method uses a hybrid architecture combining feedforward and recurrent neural network modules, with ConvLSTM layers to process sequences of measurement grid maps.
• ConvLSTM layers capture spatio-temporal dependencies by modeling motion across time and space, leveraging spatial context between neighboring cells.
• The network is trained in a supervised manner using automatically generated labels from preprocessed lidar data and existing algorithms.
• A recurrent skip connection architecture ensures dense, high-resolution predictions at the output layer.
• The input consists of measurement grid maps derived from lidar point clouds, where each cell contains an occupancy probability.
• The output is a dynamic occupancy grid map with per-cell occupancy probability and two-dimensional velocity components (vE, vN) for each occupied cell.

Experimental results

Research questions

• RQ1 Can a learning-based RNN approach improve velocity estimation accuracy in dynamic driving scenarios compared to particle filter-based methods?
• RQ2 Does the proposed method reduce inconsistent velocity estimates across neighboring cells representing the same object?
• RQ3 How does the model perform in complex maneuvers such as braking and turning, where constant velocity assumptions fail?
• RQ4 Can the network generalize to unseen urban environments without explicit object detection or data association?
• RQ5 To what extent does the model reduce erroneous velocity predictions in static areas?

Key findings

• In the stop-and-go scenario, the proposed method achieved a mean absolute error (MAE) of 0.742 m/s for velocity magnitude, outperforming the particle filter baseline (1.268 m/s).
• For velocity orientation, the proposed method achieved a mean absolute error of 8.447°, significantly lower than the baseline’s 10.515°.
• In the circular driving scenario, the proposed method reduced orientation MAE to 6.205° compared to the baseline’s 21.084°, demonstrating superior performance in curved motion.
• The proposed method achieved lower mean standard deviation (¯σvel = 0.264 m/s) in the stop-and-go scenario, indicating more consistent velocity estimates.
• Qualitative evaluation showed fewer erroneous velocity estimates in static areas, with less color noise at static obstacle contours and more tangent-aligned velocity arrows.
• The model demonstrated strong generalization to unseen urban intersections, producing velocity estimates that closely match ground truth and outperform the particle filter in dynamic scenarios.