Skip to main content

Table 3 Summary of recent SLAMMOT systems

From: A survey: which features are required for dynamic visual simultaneous localization and mapping?

References

System properties

Implementation details

Practical consideration

CT

Env

ON

OMT

MK

MMS

HD

HE

OM

HMD

SR

NP

PD

DR

Low-level based SLAM (SLAMMOT section)

 Point-based SLAM

  Wang et al. [44]

S

I

M

R

–

SSC

–

–

J

–

I

√

–

√

  Judd et al. [45]

S

I

M

R

–

MMF

–

–

J

–

I

√

–

–

Use high-level features in low-level feature-based SLAM (Using high-level features in point-based SLAM section)

 Point-based SLAM

  Nair et al. [46]

M

O

M

R

C, O

SI

L

S [27]

J

–

√

–

–

–

  Huang et al. [47]

S

I, O

M

R

–

SI

L

O [43]

S

√

I

√

√

–

  Bescos et al. [48]

S D

O

M

R

–

SI

L

S

J

√

I

√

–

–

  Ballester et al. [49]

D

O

M

R

–

SI

L

S [50]

J

√

I

√

–

–

  Zhang et al. [51]

M, S, D

I, O

M

R

–

SI

L

S [27]

J

√

-1

√

–

–

Using high-level features in object SLAM (Using high-level features in object SLAM section)

 Yang and Scherer [14]

M

I, O

M

R

–

SI

L

O [43]

S

–

–

√

–

–

 Qiu et al. [52]

M

I

S2

R

C3

SI

NN [53]

O [54]

S

–

√

√

–

–

 Strecke et al. [55]

D

I

M

R

–

SI

L

S [27]

 

√

I

√

√

√

  1. System properties: Camera type (CT): RGB-D (D), monocular (M), stereo (S). Environment (Env): indoor (I), outdoor (O). Object number (ON): single (S), multiple (M). Object motion type (OMT): rigid (R), non-rigid (NR), motion knowledge (MK): need knowledge about regarding object motion (O), need knowledge regarding camera motion (C), need no knowledge regarding motion (−). Details: Multi-motion segmentation (MMS): sub-space cluster (SSC), multi-motion fitting (MMF), semantic information (SI). High-level data association for object SLAM (HD) low-level-feature-based method (L), neural-network-based method (NN). High-level feature extractor and for object SLAM (HE): semantic segmentation network (S), object detection network (O). Optimization method (OM): joint optimization (J), separate optimization (S). Practical Consideration: Handle missing data (e.g., due to occlusion, lost tracks, motion blur) (HMD). Solve the relative-scale problem (SR): irrelevant for the type of camera (I). No need for shape priors (NP). Probabilistic data association (PD). Dense reconstruction (DR). 1. Cannot solve the relative-scale problem of monocular cameras; 2. Can implement MOT using multi-region BA; 3. Camera motion information comes from the IMU