基于时空信息融合的无人艇水面目标检测跟踪

在无人艇(USV)的导航、避障等多种任务中，目标检测与跟踪都十分重要，但水面环境复杂，存在目标尺度变化、遮挡、光照变化以及摄像头抖动等诸多问题。该文提出基于时空信息融合的无人艇水面视觉目标检测跟踪，在空间上利用深度学习检测，提取单帧深度语义特征，在时间上利用相关滤波跟踪，计算帧间方向梯度特征相关性，通过特征对比将时空信息进行融合，实现了持续稳定地对水面目标进行检测与跟踪，兼顾了实时性和鲁棒性。实验结果表明，该算法平均检测速度和精度相对较高，在检测跟踪速度为15 fps情况下，检测跟踪精确度为0.83。

Object Detection and Tracking of Unmanned Surface Vehicles Based on Spatial-temporal Information Fusion

Object detection and tracking is essential in the navigation, obstacle avoidance and other tasks of Unmanned Surface Vehicles (USV). However, the environment on the water is complex, and there are many problems such as object scale variation, occlusion, illumination variation and camera shaking, etc. This paper proposes the visual object detection and tracking of USV based on spatial-temporal information fusion. Deep learning detection in space is used to extract single-frame depth semantic features and correlation filter tracking in time is used to calculate the correlation of oriented gradient feature between frames. Temporal and spatial information through feature comparison are combined to achieve continuous and stable object detection and tracking with strong robustness at real-time. The experiments results demonstrate that the average detection and tracking accuracy is 0.83 with the average running speed of 15 fps, which illustrates the accuracy is improved and the speed is high.