Presently, although many impressed SLAM systems have achieved exceptional accuracy in a real environment, most of them are verified in the static environment. However, for mobile robots and autonomous driving, the dynamic objects in the scene can result in tracking failure or large deviation during pose estimation. In this paper, a general visual SLAM system for dynamic scenes with multiple sensors called DMS-SLAM is proposed. First, the combination of GMS and sliding window is used to achieve the initialization of the system, which can eliminate the influence of dynamic objects and construct a static initialization 3D map. Then, the corresponding 3D points of the current frame in the local map are obtained by reprojection. These points are combined with the constant speed model or reference frame model to achieve the position estimation of the current frame and the update of the 3D map points in the local map. Finally, the keyframes selected by the tracking module are combined with the GMS feature matching algorithm to add static 3D map points to the local map. DMS-SLAM implements pose tracking, closed-loop detection and relocalization based on static 3D map points of the local map and supports monocular, stereo and RGB-D visual sensors in dynamic scenes. Exhaustive evaluation in public TUM and KITTI datasets demonstrates that DMS-SLAM outperforms state-of-the-art visual SLAM systems in accuracy and speed in dynamic scenes.

In order to address the problems of various interference factors and small sample acquisition in surface floating object detection, an object detection algorithm combining spatial and frequency domains is proposed. Firstly, a rough texture detection is performed in a spatial domain. A Fused Histogram of Oriented Gradient (FHOG) is combined with a Gray Level Co-occurrence Matrix (GLCM) to describe global and local information of floating objects, and sliding windows are classified by Support Vector Machines (SVM) with new texture features. Then, a novel frequency-based saliency detection method used in complex scenes is proposed. It adopts global and local low-rank decompositions to remove redundant regions caused by multiple interferences and retain floating objects. The final detection result is obtained by a strategy of combining bounding boxes from different processing domains. Experimental results show that the overall performance of the proposed method is superior to other popular methods, including traditional image segmentation, saliency detection, hand-crafted texture detection, and Convolutional Neural Network Based (CNN-based) object detection. The proposed method is characterized by small sample training and strong anti-interference ability in complex water scenes like ripple, reflection, and uneven illumination. The average precision of the proposed is 97.2%, with only 0.504 seconds of time consumption.

In order to remove the strong noise with complex shapes and high density in nuclear radiation scenes, a lightweight network composed of a Noise Learning Unit (NLU) and Texture Learning Unit (TLU) was designed. The NLU is bilinearly composed of a Multi-scale Kernel Module (MKM) and a Residual Module (RM), which learn non-local information and high-level features, respectively. Both the MKM and RM have receptive field blocks and attention blocks to enlarge receptive fields and enhance features. The TLU is at the bottom of the NLU and learns textures through an independent loss. The entire network adopts a Mish activation function and asymmetric convolutions to improve the overall performance. Compared with 12 denoising methods on our nuclear radiation dataset, the proposed method has the fewest model parameters, the highest quantitative metrics, and the best perceptual satisfaction, indicating its high denoising efficiency and rich texture retention.