2.1. ORB-SLAM2

The ORB-SLAM2 system was first proposed by Mur-Artal et al. [Reference Mur-Artal and Tardos7], which is also a representative work using the front-end visual odometry feature point method. The system constructs three multi-thread frameworks that can run in real time. The framework has also been employed and improved by many scholars on the application of VSLAM systems. Figure 1 depicts the ORB-SLAM2 system framework.

Figure 1. ORB-SLAM2 system framework.

The tracking performs ORB feature extraction on each input frame image and performs interframe feature point matching. The local mapping is responsible for managing the keyframe obtained. The common view relationship with other keyframe is determined according to the map point information contained in the keyframe. Thus, the unqualified map points and redundant keyframe can be eliminated. Finally, a series of keyframe groups with common view relationship and their observed map points in a sliding window are optimized by a local BA. The loop closing starts to work when it detects that the process of loop closing is about to occur. It uses loop closing constraints to fuse keyframe and optimize graphs to eliminate the cumulative error generated by the system during operation.

2.2. The framework of Sem-VSLAM

Based on the framework of ORB-SLAM2, the Sem-VSLAM improves the front-end visual odometry part by adding parallel threads of semantic segmentation, detection and restoration, adding motion consistency detection algorithm in the tracking thread, adding a 3D point cloud map thread. The purpose of such improvement is to detect the dynamic object to reduce the interference of dynamic feature points to the visual odometry. In Figure 2, the Sem-VSLAM system framework is displayed.

Figure 2. The system framework of sem-VSLAM.

The improved system mainly includes four threads that can run in parallel in real time. When each RGB-D frame is input, it is transmitted to the tracking and the semantic, and the RGB-D image is processed in parallel. The Mask R-CNN [Reference He, Gkioxari, Dollár and Girshick23] network is utilized by the semantic, which divides objects into static objects and dynamic objects. The function of detection and restoration starts to repair if a lack of segmentation is detected. Then, the pixel-level semantic labels of dynamic objects are provided to the tracking thread, and the geometric constraint of motion consistency is used to further detect outliers (abnormal values) of potential dynamic feature points. In this way, the feature points on the dynamic object are eliminated, and the remaining relatively stable static feature points are used to estimate the poses. Finally, a 3D point cloud map can be constructed by combining semantic information and tracking information.

2.3. Mask R-CNN

To solve the problem of low accuracy and poor robustness of VSLAM in dynamic environment, this article combines a deep learning method to detect dynamic objects. The dynamic objects are identified using a semantic segmentation network and pixel-level semantic segmentation of dynamic objects are obtained as semantic prior knowledge. The article uses the Mask R-CNN [Reference He, Gkioxari, Dollár and Girshick23] network to perform the semantic segmentation. The network can not only obtain pixel-level semantic labels but also obtain instance labels. This article mainly uses pixel-level semantic tags to detect dynamic objects, while instance tags can also be useful in the future tracking different dynamic objects.

The Mask R-CNN is a framework that extends the Faster R-CNN [Reference Ren, He, Girshick and Sun27] object detection framework to instance segmentation. The prior information of its motion properties can be further gained using the pixel-level semantic labels that the Mask R-CNN network has obtained. For example, if the pixel-level label is “human,” then it is assumed that the pixels are a dynamic object with high confidence, because according to people’s common sense, people tend to move; if the pixel-level label is “desk,” then it is assumed that the pixels are a static object; if the pixel-level label is a “chair,” then it cannot be assumed that the pixel is a static object with high confidence. Because the chair itself cannot move, but there may be moving behavior under human activity, the pixel of this object is defined as a potentially dynamic object.

To perform semantic segmentation in the indoor environment, this article uses the Mask R-CNN under the TensorFlow framework. The pretraining model of the network is fully trained through the MS COCO dataset [Reference Lin, Maire, Belongie, Hays, Perona, Ramanan and Zitnick28], which has a high recognition and segmentation effect. The MS COCO dataset has a total of more than 80 different categories of objects. One is moving objects with high dynamic confidence, and 18 object categories in the MS COCO dataset are selected (e.g., people, bicycles, cars, motorcycles, aircraft cats, birds, and dogs); the other is the potential dynamic objects that move as people move, mainly selecting three object categories in the MS COCO dataset (e.g., chairs, books, and cups). If there are special requirements for the new categories, the network can also be retrained. The trained neural network is trained using the TUM in Germany, which can avoid spending a lot of time of collecting image data, and the plenty of data can also guarantee to obtain a robust result.

2.4. Motion consistency detection

Most dynamic objects can be segmented using the Mask R-CNN network utilized in this article, but the segmentation effect for potential dynamic objects that may move is not satisfying for the realistic applications. For example, books carried by people and chairs moved with people. To overcome this difficulty, using multi-view epipolar geometric constraints to further detect whether the features of objects are dynamic features. If the features of objects are dynamic features, they cannot meet the multi-view epipolar geometric constraints. If they are static features, they meet the multi-view epipolar geometric constraints.

Figure 3. An illustration of multi-view epipolar geometric constraints.

Figure 3(a) depicts the relationship between two successive frame image points, where $\boldsymbol{P}$ is a point in space, $\boldsymbol{P}_{1}$ and $\boldsymbol{P}_{2}$ are obtained from two consecutive frame images $\boldsymbol{I}_{1}$ and $\boldsymbol{I}_{2}$ , respectively. The baseline is defined as $\boldsymbol{O}_{1}$ and $\boldsymbol{O}_{2}$ . The epipolar plane is a plane $\pi$ that is determined by the space point $\boldsymbol{P}$ and the baseline. The intersection lines $\boldsymbol{l}_{1}$ and $\boldsymbol{l}_{2}$ of plane $\boldsymbol{I}_{1}$ and $\boldsymbol{I}_{2}$ with plane $\pi$ are called polar lines. The intersection points $\boldsymbol{E}_{1}$ and $\boldsymbol{E}_{2}$ of the image plane $\boldsymbol{I}_{1}$ and $\boldsymbol{I}_{2}$ with the baseline are called the poles. The homogeneous coordinates of feature points $\boldsymbol{P}_{1}$ and $\boldsymbol{P}_{2}$ can be expressed as

(1) \begin{align} \boldsymbol{P}_{1}=\left[\boldsymbol{u}_{1},\,\boldsymbol{v}_{1},1\right]^{\mathrm{T}},\nonumber\\[5pt] \boldsymbol{P}_{2}=\left[\boldsymbol{u}_{2},\,\boldsymbol{v}_{2},1\right]^{\mathrm{T}}, \end{align}

where $\boldsymbol{u}_{\boldsymbol{i}}, \boldsymbol{v}_{\boldsymbol{i}}(\boldsymbol{i}=1,2)$ are coordinates of $\boldsymbol{P}_{\boldsymbol{i}}(\boldsymbol{i}=1,2)$ , respectively. The polar line $\boldsymbol{l}_{1}$ is

(2) \begin{equation} \boldsymbol{l}_{1}=\left[\begin{array}{l} \boldsymbol{X}\\[2pt] \boldsymbol{Y}\\[2pt] \boldsymbol{Z} \end{array}\right]=\boldsymbol{F}\boldsymbol{P}_{1}=\boldsymbol{F}\left[\begin{array}{l} \boldsymbol{u}_{1}\\[2pt] \boldsymbol{v}_{1}\\[2pt] 1 \end{array}\right], \end{equation}

where $\boldsymbol{F}$ is a fundamental matrix, which transform the features point to the polar line. The mapping relationship can be expressed as

(3) \begin{equation} \boldsymbol{P}_{2}^{\mathrm{T}}\boldsymbol{F}\boldsymbol{P}_{1}=0. \end{equation}

Under the premise that point $\boldsymbol{P}_{1}$ in $\boldsymbol{I}_{1}$ and the fundamental matrix $\boldsymbol{F}$ are known, if $\boldsymbol{P}$ is a static point, it must satisfy the constraint of (3). However, due to the uncertainty of feature extraction and fundamental matrix $\boldsymbol{F}$ estimation, the feature points located near the epipolar line usually have errors, resulting in two image points of spatial point mapping not satisfying the constraints of (3). Feature point $\boldsymbol{P}_{2}$ is quite close to epipolar line $\boldsymbol{l}_{2}$ , as shown in Figure 3(b). Therefore, judgment criteria need to be added to evaluate the degree of the mismatch cases. A distance $\boldsymbol{D}$ from $\boldsymbol{P}_{2}$ to $\boldsymbol{l}_{2}$ can be determined as

(4) \begin{equation} \boldsymbol{D}=\frac{\left| \boldsymbol{P}_{2}^{\mathrm{T}}\boldsymbol{F}\boldsymbol{P}_{1}\right| }{\sqrt{\left\| \boldsymbol{X}\right\| ^{2}+\left\| \boldsymbol{Y}\right\| ^{2}}}. \end{equation}

By calculating (4), if $\boldsymbol{D}$ is less than a preset threshold, point $\boldsymbol{P}$ is viewed to be static; otherwise, point $\boldsymbol{P}$ is viewed to be dynamic.

Figure 4 depicts the flow chart for the method used to detect motion consistency. First, according to the previous frame feature point set $\boldsymbol{L}_{1}$ , the feature point set $\boldsymbol{L}_{2}$ in the current frame is determined using the optical flow method. Then the fundamental matrix $\boldsymbol{F}$ is estimated using at least 5 pairs of feature matching pairings, where the classical eight-point method is usually used. Finally, the relationship between the distance of the corresponding epipolar line from $\boldsymbol{P}_{2}$ to $\boldsymbol{P}_{1}$ and the preset threshold is evaluated to ascertain whether the feature point is moving or not. If it moves, it is a dynamic feature, otherwise it is a static feature.

Figure 4. The flowchart of the motion consistency detection algorithm.

2.5. Detection and restoration

In this article, besides introducing Mask R-CNN to ORB-SLAM2, and the network is also lightweighted to ensure meeting real-time requirements. However, in the lightweight processing of this network, the network emerges unstable segmentation quality, especially when the human limbs in the image are not displayed completely or the image is blurred, the segmentation quality is seriously reduced, which can result in a decline in the accuracy of pose estimation and mapping in the follow-up process. This article provides a corresponding measure to repair the segmentation results to address this issue. Each segmentation result is qualified based on the assumption that the camera has a high frame rate. If it is judged to be qualified, it is used directly; if it is judged to be unqualified, corresponding measures are taken to repair it and then use it again.

In the indoor environment, people are usually considered as dynamic objects with high dynamic characteristics. Therefore, this article mainly detects the segmentation results with lackness of segmentation for human. Let $\boldsymbol{W}_{\boldsymbol{i}}$ be the current segmentation result and $\boldsymbol{n}_{\boldsymbol{d}}(\boldsymbol{W}_{\boldsymbol{i}})$ be the number of points in the current segmentation result. To better reflect the $\boldsymbol{n}_{\boldsymbol{d}}(\boldsymbol{W}_{\boldsymbol{i}})$ close to the real change in a short time, the images that meet the segmentation qualification requirement are selected, and the serial number of the segmented image is close to the current segmented image as much as possible. The change of the people number $\boldsymbol{W}_{\boldsymbol{i}-\boldsymbol{k}}$ can be denoted as $\Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}=\boldsymbol{n}_{\boldsymbol{d}}(\boldsymbol{W}_{\boldsymbol{i}})-\boldsymbol{n}_{\boldsymbol{d}}(\boldsymbol{W}_{\boldsymbol{i}-\boldsymbol{k}})$ . If $\Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}\ll 0$ , it may be due to the lack of segmentation of $\boldsymbol{W}_{\boldsymbol{i}}$ , which is probably caused by the excessive speed of human movement or the excessive angular velocity of human from the front to the side. In both of the two cases (the excessive speed and the excessive angular velocity), the probability of $\boldsymbol{W}_{\boldsymbol{i}}$ missing segmentation is even higher than the single case. Thus, to increase the accuracy with which the absence of segmentation is detected, $\Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}\ll 0$ is judged as a sign of lack of segmentation.

Considering that the size of $\Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}$ requires a reference value to measure, this reference value requires to be able to change with the speed of the camera’s movement and also have the characteristics of being less affected by noise such as over-segmentation and lack of segmentation. Therefore, a new concept dynamic reference variation $\Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}}$ is introduced to reflect the size of $\Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}$ , and an array $\boldsymbol{Q}$ is used to store a specific $| \Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}|$ to calculate $\Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}+1}$ . The initial value of $\Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}+1}$ is set to 500.

According to the size of $\Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}$ relative to $\Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}}$ , the states of the segmentation result are divided into three categories: qualified, lack of segmentation (slight), and lack of segmentation (serious). As shown in Figure 5, three criteria are listed for judging the states of the segmentation results. Here, $\Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}}$ is the reference standard for the current segmentation results, $\boldsymbol{N}$ is a hyperparameter (set to 5), and $\alpha _{\boldsymbol{k}}, \beta _{\boldsymbol{k}}$ , and $\gamma _{\boldsymbol{k}}$ are preset thresholds and satisfy the relation: $\alpha _{\boldsymbol{k}}\lt \beta _{\boldsymbol{k}}\lt \gamma _{\boldsymbol{k}}$ .

Figure 5. Judgment criteria of segmentation result states.

After obtaining the state of $\boldsymbol{W}_{\boldsymbol{i}}$ , if $\boldsymbol{W}_{\boldsymbol{i}}$ is qualified and satisfies $0.05\times \Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}}\leq | \Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}| \leq \alpha \cdot \Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}}$ , the reference variation is updated, otherwise not updated and at the same time $\Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}+1}=\Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}}$ . The update of the reference variation is as follows:

(5) \begin{equation} \Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}+1}=\left\{\begin{array}{l} \dfrac{\boldsymbol{L}_{\boldsymbol{i}}\cdot \Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}}+\left| \Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}\right| }{\boldsymbol{L}_{\boldsymbol{i}}+1},\,\,\,\,\,\,\,\,\,\enspace \boldsymbol{L}_{\boldsymbol{i}}\lt \boldsymbol{L}_{\max }\\[16pt] \dfrac{\boldsymbol{L}_{\max }\cdot \Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}}-\boldsymbol{q}_{\boldsymbol{f}}+\left| \Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}\right| }{\boldsymbol{L}_{\max }},\boldsymbol{L}_{\boldsymbol{i}}=\boldsymbol{L}_{\max } \end{array}\right., \end{equation}

where $\boldsymbol{q}_{\boldsymbol{f}}$ is the starting element of array $\boldsymbol{Q}$ when generating $\boldsymbol{W}_{\boldsymbol{i}}, \boldsymbol{L}_{\boldsymbol{i}}$ is the $\boldsymbol{i}\mathrm{th}$ element of array $\boldsymbol{Q}$ when generating $\boldsymbol{W}_{\boldsymbol{i}}, \boldsymbol{L}_{\max }$ is a hyperparameter representing the maximum length of $\boldsymbol{Q}$ , and $\boldsymbol{L}_{\max }$ is set to 20. After updating the reference variable, $| \Delta \boldsymbol{n}_{\boldsymbol{d}}^{\boldsymbol{i}}|$ is inserted into array $\boldsymbol{Q}$ ; if $\boldsymbol{Q}\gt \boldsymbol{L}_{\max }$ , the starting element need to be discarded.

After the calculation of $\Delta \boldsymbol{n}_{\boldsymbol{ref}}^{\boldsymbol{i}+1}$ , the results lack of segmentation (slight and serious) can be repaired. Because the difference of dynamic objects in an image does not change much in a short time, the repair method adopted is to superimpose the occlusion and mask of the dynamic object by the segmentation result, which is the largest number of human points in the first $\boldsymbol{N}$ segmentation results of the current frame, on the repaired object.