From: Review of visual odometry: types, approaches, challenges, and applications
Reference | Camera type | Approach | VO estimation accuracy | Limitations |
---|---|---|---|---|
Gonzalez et al. (2013) | Two monocular cameras: downward-facing camera for displacement and front-facing camera for orientation estimation | Appearance-based approach (NCC template matching) | Error <3% of the total travelling distance and <8° average orientation error | False matches due to shadows and blur at velocity >1.5 m/s Can’t deal with scale variance on non-smooth surfaces |
Van Hamme et al. (2015) | Monocular camera | Feature-based approach (inverse perspective projection and Kalman filter for Tracking of features in the ground plane) | >8.5% translation error (for 800Â m) | Significant rotational bias on some estimated trajectory segments due to non-planarity of the road environment in those segments |
Scaramuzza and Siegwart (2008a) | Omnidirectional camera | Hybrid approach (tracking SIFT feature points from ground plane to estimate translation. Image appearance similarity measure (NCC, Manhattan and Euclidean distance) was used to estimate the rotation of the car) | Error is <2% of the distance traversed 5° average orientation error | Unavoidable visual odometry drift and deviation due to road humps that violate the planar motion assumption |
Nistér et al. (2004) | Stereo camera | Feature-based approach (Detection of features independently in all frames and only allowed matches between salient features) | 1.63% error over 380 m of the distance traversed | No mention |
Howard (2008) | Stereo camera | Feature-based approach (Feature matching and employing stereo range data for inlier detection) | 0.25% error over 400Â m of the distance traversed | Self-Shadow leads to false-matches It does not work effectively on vegetated environment |
Nourani-Vatani and Borges (2011) | Monocular camera | Appearance-based approach (NCC multi-template matching which selects best template based on entropy) | Error <5% of total travelling distance 5° average heading error | Deficiency in dealing with scale variance at uneven surfaces System can’t deal with sunny/shadow regions |
Yu et al. (2011) | Monocular camera | Appearance-based approach (NCC rotated template matching) | 1.38% distance error and 2.8° heading error | Cannot deal with image scale variance, shadows and blur |
Nagatani et al. (2010) | Telecentric camera (which maintains the same field of ground area view, regardless of variation in camera height from ground | Appearance-based approach (cross correlation template matching) | <3% error indoor experiment 1.5% (for 100Â m trajectory) at 0.4Â m/s speed | Cannot estimate the camera height from ground variations |
Zhang et al. (2014) | Monocular camera | Feature-based approach [tracking of features using Lucas Kanade Tomasi (LKT)] | Error is <1% of the distance traversed | Image scale uncertainty at complicated ground conditions for example loose soil floors |