Visual Odometry in Smoke Occluded Environments

Abstract

Visual odometry is an important sensor modality for robot control and navigation in environments when no external reference system, e.g. GPS, is available. Especially micro aerial vehicles operating in cluttered indoor environments need pose updates at high rates for position control. At the same time they are only capable to carry sensors and processors with limited weight and power consumption. Thus a need arises to compare and modify state-of-the-art methods so that the appropriate one can be identified. When these MAVs are used for as tools for inspection and damage assessment, the need to navigate in challenging and degraded indoor environments becomes essential. The work addresses the problem of odometry failure in unfavourable conditions of fire and smoke. The reasons for odometry failure are identified and various image enhancement techniques are implemented and compared. The case of contrast enhancement using image depth maps is inspected closely in particular since 3D depth data is available for use. Apart from visually enhancing the hazy image, the method also shows improvement in feature extraction, feature matching and inlier detection, all of which are essential components of visual odometry methods. Two visual odometry methods SVO to Fovis are then compared using various benchmarking and evaluation methods with the purpose of determining the more efficient and accurate method.