基于深度可分离卷积的多目标追踪神经网络研究

Multi-Target Tracking Neural Network based on Depthwise Separable Convolutions

提出一种在线的多目标追踪方法,通过指定所需追踪的目标,结合人工智能强大特征提取能力,可以实时的追踪该目标的位置.在传统神经网络的追踪算法中,引入了深度可分离卷积,利用其强大时序特征提取能力,使该网络在各种不同光照、不同角度的情况下,推理前后数帧之间的关系,得到稳定的追踪区域.除此之外,利用深度可分离卷积与改进后的卷积层相连接,使其拥有了多目标追踪能力的同时,依然满足实时处理能力.在公开数据机上VOT-2016测试结果显示,与现有追踪方法相比,该方法不仅在准确度,EAO的表现出色,而且保持了不错的处理速度.通过该方法可以有效地统计地铁站OD流,或追踪指定人物的移动轨迹等,满足在交通以及安防领域的需求.

An online multi-target tracking method is proposed.By specifying the target to be tracked,the powerful feature extraction capability of artificial intelligence is combined,and the position of the target can be tracked in real time.Due to the powerful feature extraction ability of the separable depth convolution,the tracking of different angles of light is very stable.Based on the SimaFC network,introducing deep separable convolution and improving the convolution,this method can track multiple features while maintaining a certain processing speed.The test results at VOT-2016 show that,compared with the existing tracking method,this method not only excels in accuracy,EAO performance,but also maintains a good processing speed.This method has good performance in tracking the OD flow statistics of subway stations,or the movement trajectories of specified characters.