**6. Discussion**

Multi-sensor fusion positioning technology based on SLAM provides new opportunities for the high-precision positioning of mobile carriers. In this study, two problems that need to be further explored in the existing LVIO fusion system are proposed. The first problem is that LVIO system needs enough environmental feature information. According to the previous studies of Pumarola et al. [11] and Fu et al. [12], theoretically, the accuracy of the fusion system can be improved by increasing the constraint of visual line features. Huang et al. also proved that the average positioning error of the fusion system based on point-line feature can generally be improved, from the traditional 2.16% to 0.93% [18]. In this study, the steps of increasing visual line feature constraints are further optimized. The Monte Carlo method was used to select the appropriate scaling ratio and the density threshold of homogeneous points, which improves the angle tolerance of pixel fitting to line features. To a certain extent, it reduces the probability that short segments are wrongly judged as invalid features. According to the experiment of parameter selection in Section 3.1, compared with the indoor environment where the angle and translation of line features change little, the movement of line features between consecutive frames is more complicated in the outdoor environment. Therefore, the density threshold of homogeneous points needs to be lowered to reduce the probability that the valid line feature pairs are misjudged as invalid matches when turning sections. The results of outdoor real-time analysis shown in Section 5.1.2. show that the traditional method based on point-line features is difficult to match and track the visual line features in the outdoor environment, which takes a long time. However, the time consumption of this algorithm maintained a low level, which is beneficial to leave more time for back-end fusion optimization.

To solve the problem of line feature mismatching during the movement of carriers, Zhou et al. established a constraint equation using the 6-DOF Plücker coordinates of line features to perform matching optimization [19]. However, this increases the computational complexity of the fusion system, which is inconsistent with the lightweight requirements of autonomous driving positioning. In this study, the link of line feature constraint matching is simplified, and the original 6-DOF parameters are replaced by 4-DOF which represents the movement of line features for optimization. Thus, it can reduce the computational complexity of the system and effectively improve the inter-frame matching accuracy of line features. To explore the superiority of the proposed algorithm in real-time and positioning accuracy, we compared the precision and the time-consuming of three processes related to line features of this algorithm with several similar advanced algorithms in different environments. The experimental results show that our optimization strategy based on front-end point-line features effectively achieves the positive balance between reducing time consumption and improving accuracy.

The second problem to be solved is the global optimization of LVIO local pose estimation results by introducing global constraints. To further improve the positioning accuracy of local sensors, Qin et al. proposed a GNSS and local sensor fusion method to construct GNSS residual factors to correct the cumulative error of VIO [25]. Further, we propose a factor graph based on Bayesian network, in which GNSS observations are added as global constraint factors. The accumulated errors of LVIO are corrected by using GNSS observations within 0.1 s interval from LVIO keyframes as global constraint. In this study, it is proved that GNSS global constraint factor can effectively correct LVIO positioning error in the outdoor environment. It should be noted that since the coordinate of the current frame of LVIO is calculated from the coordinate of the previous frame, longterm observation or long moving distance will lead to more serious data drift. However, the GNSS observations are in the global coordinate, so long-term observation is not related to the data drift. Therefore, we can reasonably speculate that the longer the algorithm runs, the more obvious the correction effect of GNSS on LVIO will be. More comprehensively, LVIO will continue local positioning in GNSS rejection environment, so the positioning continuity of mobile carriers in different environments can be effectively guaranteed.

#### **7. Conclusions**

In this study, a LiDAR-Visual-Inertial Odometry based on optimized visual point-line features is proposed, taking advantage of the heterogeneous complementary characteristics of multiple sensors. First, a visual line feature extraction and matching optimization method is proposed. By improving the line feature extraction in the scale space and selecting the appropriate scaling ratio and same-sex point density threshold, the number of line features extracted in the light complex environment is largely improved to provide richer feature information for the front-end. Meanwhile, the original 6-DOF parameter optimization problem is further improved to a 4-DOF parameter optimization problem by using a least squares-based line feature constrained matching strategy. The complexity of the fusion system is reduced, and more accurate visual pose estimation is effectively accomplished. Second, the LiDAR point cloud is projected into the visual coordinates for depth correlation. Meanwhile, the initial pose estimation provided by the optimized VIO is used to help LiDAR scan matching. Finally, a factor graph method based on Bayesian networks is established. Two global constraint factors are added to the factor graph framework to constrain LVIO globally, which are the global constraint of GNSS factors from external sensors and the loop factor constraint of local sensors. The experimental results show that the algorithm can achieve real-time attitude estimation with good localization and mapping accuracy in different environments.

In the future, we will further improve and refine our work in the following aspects. First, the point cloud alignment algorithm of the loop factor in this study utilizes the

traditional ICP algorithm, which is time-consuming to perform the nearest domain search using the KD tree. Thus, we will consider the improvement of the point cloud alignment algorithm next. Second, the inclusion of the GNSS factor in this study only utilizes the GNSS pseudo range single point positioning result. Although it is relatively simple and feasible on the vehicle platform with only one GNSS receiver, there is still room for improvement in the positioning accuracy of GNSS. A more accurate correction of LVIO by using higher accuracy RTK positioning results will be considered in the next step. Finally, since our proposed fusion system consists of two subsystems with high runtime computational resource requirements, we will work on reducing the resource occupation rate of the algorithm. Further, we will evaluate the positioning accuracy of the algorithm on vehicles with limited computing resources.

**Author Contributions:** Conceptualization, Z.Z., C.S., H.Z. and X.L.; methodology, X.H.; software, X.H.; validation, S.P. and L.D.; formal analysis, X.H.; investigation, W.G. and X.H.; resources, S.P., W.G. and X.H.; writing—original draft preparation, X.H.; writing—review and editing, W.G. and X.H.; supervision, S.P. and W.G.; project administration, W.G.; funding acquisition, S.P. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research study was funded by the Fundamental Research Funds for the Central Universities (2242021R41134) and the Research Fund of the Ministry of Education of China and China Mobile (MCM20200J01).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.
