Markov Localisation using Heatmap Regression and Deep Convolutional Odometry

Oscar Mendez Maldonado, Simon Hadfield, Richard Bowden

Abstract

This paper presents a novel approach to visual localization by integrating classical Markov localization into a deep learning framework. By reformulating grid-based localization as a series of convolutional operations, the authors achieve a robust, multi-modal system capable of handling ambiguous environments. The proposed architecture includes a deep likelihood heatmap regressor and a convolutional odometry layer, enabling real-time, GPU-accelerated probabilistic filtering. Evaluations on the Cambridge Landmarks and a new Multi-Storey Carpark dataset demonstrate superior accuracy and trajectory smoothness compared to standard pose-regression methods.

Introduction

Visual localization is a fundamental task for autonomous systems, yet common deep learning approaches like direct pose regression (e.g., PoseNet) often struggle with overconfidence and uni-modality. These models typically provide a single point estimate, failing to represent uncertainty in self-similar environments like parking garages or uniform hallways where multiple locations may look identical.

Classical Markov localization addresses these issues by maintaining a probability distribution over a grid of possible states. However, the computational cost of updating large 3D grids ($x, y, \theta$) has historically limited its application on standard CPUs. This work bridges the gap by implementing the entire Markov localization pipeline within a Convolutional Neural Network (CNN), leveraging GPU acceleration to make grid-based filtering efficient and scalable.

Methodology

The proposed system consists of three primary components integrated into a single forward pass:

  1. Deep Likelihood Heatmap Regressor: Utilizing a ResNet-50 encoder and a custom decoder, the network predicts a "birds-eye-view" likelihood map. Localization is treated as a classification problem over discrete spatial bins, allowing the model to represent multiple hypotheses simultaneously.
  2. Convolutional Odometry Layer: The motion model update is formulated as a 2D convolution. By representing robot motion as a kernel, the probability distribution can be propagated across the grid using standard deep learning operations, significantly reducing computational overhead (approx. 0.25ms per update).
  3. Multi-Scale Supervision: To ensure accurate and smooth probability distributions, the network is trained with heatmaps at four different resolutions, providing a coarse-to-fine refinement of the pose estimate.
Overview of the Markov Localisation pipeline, featuring the Heatmap Regressor and Convolutional Odometry components.

Experimental Setup

The architecture was evaluated on two distinct benchmarks:

  • Cambridge Landmarks: A standard outdoor visual localization dataset.
  • Multi-Storey Carpark Dataset: A challenging new dataset introduced by the authors, featuring a 6,500m trajectory across 12 floors of a parking garage. This environment is highly self-similar, providing a rigorous test for multi-modal localization capabilities.

Results and Discussion

The hybrid approach demonstrates significant improvements in both accuracy and robustness. On the Cambridge Landmarks dataset, the Heatmap Regressor outperformed various PoseNet baselines in 4 out of 6 scenarios. In the highly ambiguous Multi-Storey Carpark environment, the combined Heatmap and Odometry (H+O) system achieved an RMSE of 6.48m, significantly outperforming PoseNet (9.31m) and PoseLSTM (9.05m).

A key advantage of the Markov framework is temporal smoothing. By maintaining a prior distribution (memory of previous states), the system produces trajectories that are physically plausible and free from the erratic "jumps" common in frame-by-frame regression models when they encounter visual ambiguity.

Qualitative comparison of localization trajectories. The proposed H+O method (green dashed line) closely follows the ground truth (black) and maintains smoothness compared to other baselines.

Conclusion

By reimagining Markov localization as a convolutional process, this paper provides a robust solution for visual localization in complex environments. The system effectively handles uncertainty and multi-modality while remaining computationally efficient. Future work may involve extending the framework to 6-DoF localization and integrating it into full SLAM (Simultaneous Localization and Mapping) systems where the map itself is learned and updated dynamically.

Summary updated

Abstract

In the context of self-driving vehicles there is strong competition between approaches based on visual localisation and Light Detection And Ranging (LiDAR). While LiDAR provides important depth information, it is sparse in resolution and expensive. On the other hand, cameras are low-cost and recent developments in deep learning mean they can provide high localisation performance. However, several fundamental problems remain, particularly in the domain of uncertainty, where learning based approaches can be notoriously over-confident.Markov, or grid-based, localisation was an early solution to the localisation problem but fell out of favour due to its computational complexity. Representing the likelihood field as a grid (or volume) means there is a trade off between accuracy and memory size. Furthermore, it is necessary to perform expensive convolutions across the entire likelihood volume. Despite the benefit of simultaneously maintaining a likelihood for all possible locations, grid based approaches were superseded by more efficient particle filters and Monte Carlo sampling (MCL). However, MCL introduces its own problems e.g. particle deprivation.Recent advances in deep learning hardware allow large likelihood volumes to be stored directly on the GPU, along with the hardware necessary to efficiently perform GPU-bound 3D convolutions and this obviates many of the disadvantages of grid based methods. In this work, we present a novel CNN-based localisation approach that can leverage modern deep learning hardware. By implementing a grid-based Markov localisation approach directly on the GPU, we create a hybrid Convolutional Neural Network (CNN) that can perform image-based localisation and odometry-based likelihood propagation within a single neural network. The resulting approach is capable of outperforming direct pose regression methods as well as state-of-the-art localisation systems.
BibTeX
@inproceedings{maldonado2021markov,
  author = {Oscar Mendez Maldonado and Simon Hadfield and Richard Bowden},
  title = {{Markov Localisation using Heatmap Regression and Deep Convolutional Odometry}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2021},
  url = {https://openreview.net/forum?id=pLhu55DFJ-},
}
Figure