3.3.2 Analysis of Localization Robustness
Fig. 5(a) displays the Absolute Pose Error (APE) of the algorithm before and after improvements in scenario 1. With specific numerical data from Table 3, significant improvements can be observed in all error metrics of APE. For example, the Root Mean Square Error (RMSE) decreased from 1.7565 before improvement to 0.3959 afterwards, representing an improvement of approximately 77%. The Maximum Error (Max) also decreased from 2.7385 to 0.7843, a reduction of over 70%, indicating that the improved algorithm can more effectively control the maximum pose error. Additionally, the Mean Error decreased from 1.6543 to 0.3630, a drop of around 78%, demonstrating a significant reduction in the average error. These data suggest that incorporating the relocalization module has notably increased the global pose estimation accuracy of Lidar SLAM. This test indicates that our relocalization module can significantly reduce the positioning error in crop row scenes.
Fig. 5(b) presents the test results in a flat greenhouse environment. By observing the data in Table 3, it can be observed that the improved algorithm shows some degree of improvement in APE. For example, the RMSE decreased by approximately 20%, from 4.3098 to 3.4303. The Mean Error also reduced from 3.9928 to 3.0176, representing a decrease of around 24%. However, the level of improvement in other key metrics is relatively small. The Maximum Error only decreased by about 4%, and the Standard Deviation (Std) did not change significantly. On the other hand, a slight increase in Std may suggest that under certain conditions, such as specific noise or environmental changes, the improved algorithm may still have areas that require further improvement. From this test, it can be concluded that our relocalization module can generally reduce the positioning error in a relatively simple environment. However, the effects may not be as pronounced as in more complex environments. This could be attributed to the smaller pose changes between consecutive frames in simpler environments, where even minor errors can potentially impact the stability of APE.
Fig. 5(c) presents the test results from scenario 3. When comparing the algorithms before and after enhancement, a significant improvement in APE is evident, as shown in Table 3. Particularly, the RMSE value decreased from 4.4524 to 2.3813, representing a reduction of over 46%. These demonstrate the effectiveness of our relocation algorithm in an environment with various terrains and interference factors. From this test, it can be concluded that our relocalization module has excellent effects in reducing global positioning errors in complex scenarios.
Fig. 5(d) showcases the test results from scenario 4, which has many low-lying plants. Referring to Table 3 and comparing the algorithms before and after enhancement in terms of APE, it is evident that almost all metrics of the enhanced algorithm have significantly improved. For example, the RMSE dropped from 3.4979 to 2.3496, representing a decrease of about 33%. This indicates that our relocation module can still provide significant optimization in scenes with dirt roads and interference from low-lying plants.
These results indicate that our improved algorithm demonstrates notable efficacy across various environments. The relocalization module consistently reduces the APE in all experimental scenarios, leading to an enhancement in the accuracy of global pose estimation. This improved performance is particularly noticeable in complex environments with rugged roads or low-lying plant interference. Therefore, the Improved LeGO-LOAM algorithm in this study is capable of providing precise localization information for adaptive navigation within a greenhouse.
Table 3. Comparisons of the APE values between the Two Methods.