No abstract

As we all know, many dynamic objects appear almost continuously in the real world that are immensely capable of impairing the performance of the majority of vision-based SLAM systems based on the static-world assumption. In order to improve the robustness and accuracy of visual SLAM in high-dynamic environments, a real-time and robust stereo SLAM system for dynamic scenes was proposed. To weaken the influence of dynamic content, the moving-object detection method was put forward in our visual odometry, and then the semantic segmentation network was combined in our stereo SLAM to extract pixel-level contours of dynamic objects. Then, the influences of dynamic objects were significantly weakened and the performance of our system increased markedly in dynamic, complex, and crowed city spaces. Following experiments with both the KITTI Odometry dataset and in a real-life scene, the results showed that our method could dramatically decrease the tracking error or drift, and improve the robustness and stability of our stereo SLAM in high dynamic outdoor scenarios.

Facing the scale variation challenge in the topic of object detection, in this papers we design feature fusion methods to improve the representation ability of features. The proposed network, Feature Fusion Network (FFNet), contains mainly two parts: the Relationship Fusion Module (RFM) and the Numerical Fusion Module (NFM). The long dependencies information is shared within the pyramidal features in RFM. This information helps each feature emphasize informative regions and reduce the influence of useless regions. The NFM introduces the averaging operation to generate the fusion weights to retain the useful information and conduct the efficient feature fusion in the lateral connection. Finally, we conduct extensive experiments to verify the effects of our methods.

No abstract

No abstract