【摘要】 为有效解决传统人工标注定位车辆行为存在的检测率低且相应的目标检测算法实用性差的弊端,提出一种基于深度学习的车辆时序动作检测算法,将视频中车辆直行行为设为背景行为,车辆转向、掉头等行为设定为目标行为。利用双流卷积网络对长视频中目标行为进行提取得到初级区域提议,利用双向长短记忆网络对得到的初级提议进行细化裁剪操作,实现对车辆行为类别的检测以及该行为的时间提取。实验结果表明,该算法与其它算法进行比较,在平均精度和时间交并比上均较优。更多还原

【Abstract】 To effectively solve the disadvantages of low detection rate of traditional manual labeling and positioning of vehicle behavior and poor practicability of the corresponding target detection algorithm,a vehicle sequential action detection algorithm based on deep learning was proposed.The straight behavior of the vehicle in the video was set as the background behavior,and the behavior of the vehicle turning,U-turn,etc.were set as the target behavior.The dual-stream convolutional network was used to extract the target behavior in the long video to obtain the primary region proposal,and the two-way long and short memory network was used to refine the primary proposal obtained,and the detection of the vehicle behavior category and the time extraction of the behavior were.The experimental results of the algorithm were compared with other algorithms.The proposed algorithm is better than other algorithms in average accuracy and time intersection ratio.更多还原