Robust outdoor trajectory mapping using CNN features and loop closure optimization

Autonomous navigation in cluttered outdoor environments remains a significant challenge for robotic systems due to dynamic obstacles, occlusions, and feature degradation. Traditional Simultaneous Localization and Mapping (SLAM) systems, reliant on low-level geometric features, often fail to accurately estimate trajectory in such conditions. This paper introduces a robust framework for outdoor trajectory mapping that synergizes deep learning-based feature extraction with loop closure detection and optimization to address these limitations. To find a suitable layer to obtain features from the CNN model, 313,746 filters were checked to find a filter that has the least odometry error from three pre-trained CNN models, the ConvNeXtXLarge is leveraged to extract high-level semantic features from monocular images, enabling resilient optical flow estimation even in scenes with transient objects and lighting variations. These features are fused with inertial measurement unit (IMU) data to compute initial trajectories, minimizing scale ambiguity. To counteract accumulated drift, a hybrid optimization strategy is proposed: Bag of Words (BoW)-based loop closure detection identifies revisit locations, while an algorithm is developed to optimize the map. Experimental validation on the KITTI dataset demonstrates a significant reduction in offset error compared to state-of-the-art methods like DeepVO and TartanAir.