基于视觉的水面背景下目标检测与跟踪算法

Target Detection and Tracking Algorithm Based on Vision in Water Surface Background

为解决清漂船在复杂水面背景下对漂浮物体积较小或被遮挡的检测与跟踪问题,提出了一种基于视觉的水面背景下目标检测与跟踪算法,通过利用YOLO改进的多粒度特征融合方法使得模型在最终检测时所提取的特征向量考虑更多底层的特征,并引入K邻域搜索感兴趣区域模块,与长短期记忆神经网络(long-short term memory,LSTM)相结合,弥补了卷积神经网络的时序关联性差的缺陷,根据目标当前帧语义特征和运动特征预测下一帧中目标所在位置,能够更快地提取目标的特征,并且有效地去除复杂背景的干扰。实验结果表明:该算法的跟踪平均成功率、平均准确率和平均速度分别为57.1%、71.1%、45.4帧/s。较好解决因被检测目标过小的问题,提升在跟踪目标被遮挡的情况下的跟踪性能。

In order to solve the clean boat problem of detecting and tracking the small volume or occlusion of floater in the background of complex water surface,a target detection and tracking method under the background of surface based on visual algorithm was proposed.YOLO improved multiple granularity characteristics fusion method was used to make the model in the final test when the extracted feature vector to consider more features of the underlying.And K neighborhood search interested area module was introduced.The module combined with both short-term and long-short term memory(LSTM)made up the temporal correlation of the convolutional neural network with poor defects,according to the target current frame semantic features and motion prediction target location in the next frame,which makes the feature extraction of the target faster and effectively removes the interference of complex background.The results show that the average success rate,average accuracy and average speed of the algorithm are 57.1%,71.1%and 45.4 frame/s,respectively.It is concluded that the algorithm can better solve the problem that the detected target is too small and improve the tracking performance when the tracking target is obscured.