# Multi-Modal Obstacle Detection in Unstructured Environments with Conditional Random Fields

## TL;DR

This paper presents a multimodal obstacle detection system combining lidar and camera data with a conditional random field to improve perception in unstructured agricultural environments, addressing challenges of geometric and appearance variability.

## Contribution

It adapts a state-of-the-art multimodal fusion algorithm for obstacle detection in agriculture, explicitly handling sparse lidar data and integrating spatial, temporal, and multimodal links.

## Key findings

- Modest 0.5% improvement in two-class classification with camera data alone.
- Significant 7.9% improvement in multi-class 3D classification when combining modalities.
- Temporal links yield additional 1.5% accuracy gains in 3D classification.

## Abstract

Reliable obstacle detection and classification in rough and unstructured terrain such as agricultural fields or orchards remains a challenging problem. These environments involve large variations in both geometry and appearance, challenging perception systems that rely on only a single sensor modality. Geometrically, tall grass, fallen leaves, or terrain roughness can mistakenly be perceived as nontraversable or might even obscure actual obstacles. Likewise, traversable grass or dirt roads and obstacles such as trees and bushes might be visually ambiguous. In this paper, we combine appearance- and geometry-based detection methods by probabilistically fusing lidar and camera sensing with semantic segmentation using a conditional random field. We apply a state-of-the-art multimodal fusion algorithm from the scene analysis domain and adjust it for obstacle detection in agriculture with moving ground vehicles. This involves explicitly handling sparse point cloud data and exploiting both spatial, temporal, and multimodal links between corresponding 2D and 3D regions. The proposed method was evaluated on a diverse data set, comprising a dairy paddock and different orchards gathered with a perception research robot in Australia. Results showed that for a two-class classification problem (ground and nonground), only the camera leveraged from information provided by the other modality with an increase in the mean classification score of 0.5%. However, as more classes were introduced (ground, sky, vegetation, and object), both modalities complemented each other with improvements of 1.4% in 2D and 7.9% in 3D. Finally, introducing temporal links between successive frames resulted in improvements of 0.2% in 2D and 1.5% in 3D.