基于改进SSD模型和自适应滤波算法的水面漂浮物目标检测跟踪方法

水面漂浮物严重破坏河道景观和水生态环境，通过摄像头实施河湖可视化监管以改善河湖面貌，已成为积极落实“河湖长制”政策的重要技术手段。但由于河流环境复杂多样，存在水波扰动、动态光影和强光反射等诸多噪声问题，现有方法难以满足水面漂浮物实际管理需求。本文结合单帧检测与多帧滤波，提出了一种基于深度学习的水面漂浮物目标智能检测跟踪方法。在单帧检测中，删除5*5及以下低分辨率特征图，通过特征求和方式增强76*76高分辨率特征图以提升SSD检测算法在小尺度漂浮物目标的检测精度；在多帧滤波中，基于时空相关性和运动信息构建自适应滤波AF算法框架，计算视频帧中帧与帧之间的关联性，根据漂浮物目标的空间位置的变化幅度自适应删除偏离运动轨迹的漂浮物目标检测结果，以降低漂浮物跟踪漂移；在信息融合阶段，通过特征对比融合检测和滤波信息，实现检测信息和跟踪信息动态互补，并以不同检测跟踪场景数据集进行训练与验证。结果表明：在简单水面场景下中心位置误差为8个像素点时该方法的检测跟踪精度达到100%，成功率面积为0.94，平均速度达到17.27FPS，计算复杂度为7.18Billion；在复杂水面场景下中心位置误差为10个像素点时该方法的跟踪精度和成功率面积分别为93.24%和0.81，平均速度和计算复杂度分别为15.02FPS和8.76Billion，在复杂环境下兼顾了检测跟踪精度和效率。

Detection and Tracking Method of Floating Objects on Water Surface Based on Improved SSD Model and Adaptive Filtering Algorithm

The floating objects on the water surface seriously damage the river landscape and the water ecological environment. However, due to the complex and diverse river environment, there are many noise problems such as water wave disturbance, dynamic light and shadow, and strong light reflection. To solve the problem of floating object detection and tracking in complex environments, a deep learning-based intelligent detection and tracking method for floating object targets on the water surface is proposed by combining single-frame detection and multi-frame filtering. In single-frame detection, low-resolution feature maps of 5*5 and below are removed, and high-resolution feature maps of 76*76 are enhanced by feature added to improve the detection accuracy of the SSD detection algorithm for small-scale floating objects. In multi-frame filtering, an adaptive filtering algorithm framework is constructed based on spatial-temporal correlation and motion information to calculate the correlation between frames in a video frame and adaptively remove the detection results of floater targets that deviate from the motion trajectory according to the magnitude of change in the spatial position of the floater target in order to reduce floater tracking drift. In the information fusion stage, the detection and filter information is fused by feature comparison to achieve dynamic complementarity between detection and tracking information and are trained and validated with different detection and tracking scene datasets. The results show that the method achieves 100% detection and tracking accuracy, 0.94 success rate area, 17.27 FPS average speed, and 7.18 Billion computational complexity in simple water scenes with a center location error of 8 pixels. The tracking accuracy and success rate area for a center location error of 10 pixels in complex water scenes are 93.24% and 0.81 respectively, and the average speed and computational complexity are 15.02 FPS and 8.76 Billion respectively, balancing detection and tracking accuracy and efficiency in complex environments.