Bayesian Spatial Kernel Smoothing for Scalable Dense Semantic Mapping (1909.04631v2)

Abstract: This paper develops a Bayesian continuous 3D semantic occupancy map from noisy point clouds by generalizing the Bayesian kernel inference model for building occupancy maps, a binary problem, to semantic maps, a multi-class problem. The proposed method provides a unified probabilistic model for both occupancy and semantic probabilities and nicely reverts to the original occupancy mapping framework when only one occupied class exists in obtained measurements. The Bayesian spatial kernel inference relaxes the independent grid assumption and brings smoothness and continuity to the map inference, enabling to exploit local correlations present in the environment and increasing the performance. The accompanying software uses multi-threading and vectorization, and runs at about 2 Hz on a laptop CPU. Evaluations using multiple sequences of stereo camera and LiDAR datasets show that the proposed method consistently outperforms current baselines. We also present a qualitative evaluation using data collected with a bipedal robot platform on the University of Michigan - North Campus.

• The paper introduces a Bayesian framework using spatial kernel smoothing to create scalable, continuous multi-class 3D semantic maps from noisy point clouds.
• The proposed method consistently outperforms existing baseline techniques on challenging datasets like KITTI, achieving improved accuracy and effectively interpolating sparse measurements.
• Unlike independent grid methods, the Bayesian spatial kernel approach leverages local correlations for smoothness and continuity, demonstrated effectively on robotic platforms.

An Overview of Bayesian Spatial Kernel Smoothing for Scalable Dense Semantic Mapping

This paper presents an advancement in the field of semantic mapping by developing a Bayesian framework for continuous three-dimensional (3D) semantic occupancy maps derived from noisy point clouds. This innovative approach extends the binary Bayesian kernel inference model, traditionally used for occupancy maps, to accommodate multi-class semantic mapping. The proposed method synergistically models both occupancy and semantic probability estimates, defaulting sensibly to classic occupancy frameworks when measurements yield a single occupied class.

A notable contribution of this work is its departure from the independent grid assumption typical of conventional mapping techniques. By adopting a Bayesian spatial kernel inference, the method affords smoothness and continuity to map inference, enhancing the model's ability to leverage local environmental correlations. The software accompanying this paper, optimized via multi-threading and vectorization, demonstrates practical efficacy by operating at approximately 2 Hz on a standard laptop CPU.

The experimental evaluations underscore the efficacy of the proposed model. Notably, it consistently surpasses existing baseline methods when applied to sequences from stereo camera and LiDAR datasets. A qualitative evaluation with data gathered from a bipedal robot at the University of Michigan further exemplifies its application potential. The proposed methodology shows a robust improvement in mapping accuracy, effectively closing gaps in mapped data and interpolating sparse measurements to realize a more accurate representation of the environment.

Technical Specifics and Methodology:

The methodology introduces a Bayesian continuous inference model that accommodates multi-class data, contrasting the existing binary occupancy mapping schemes. Utilizing Bayesian Kernel Inference (BKI), this framework is applied to a Categorical likelihood with a Dirichlet distribution prior. This choice facilitates efficient computation and retains the model’s scalability. The concept of smoothing through a kernel function diverges from conventional CRF models, which typically rely on post-processing for inconsistency mitigation.

The sparse kernel function chosen is instrumental in reducing computational complexity while allowing local measurement correlations. This is particularly beneficial in contexts where point cloud data exhibits high variability in density or noise.

Performance and Results:

Quantitative results demonstrate superior outcomes in terms of Intersection over Union (IoU) metrics, with significant improvements observed, particularly on the challenging KITTI sequences. Moreover, the continuity and predictive capabilities of this semantic mapping approach are validated through comparative analysis against CRF-based systems and discrete occupancy grid approaches. Furthermore, the robustness of the model is affirmed through applications on both stereo camera and LiDAR data, along with practical deployment demonstrations on autonomous robot platforms.

Implications and Future Directions:

The implications of this research extend into multiple domains, including robotics navigation and autonomous vehicle environments, where precise and continuous semantic mapping is crucial. The work suggests promising avenues for future research, particularly in terms of automated hyperparameter tuning within a Bayesian framework and advancements in dynamic scene mapping to distinguish static from dynamic elements.

In conclusion, the paper contributes significantly to the advancement of semantic occupancy mapping by addressing key limitations in existing methodologies and demonstrating a practical, scalable solution with breadth in application potential. Future developments may explore optimization strategies for memory efficiency and dynamic mapping capabilities, paving the way for broader adoption and integration into real-world navigational systems.