空间上下文与时序特征融合的交警指挥手势识别技术

针对无人驾驶汽车快速准确识别交警指挥手势的需求，本文在分析交警指挥手势的关节铰接特征基础上，建立基于关节点和骨架的交警指挥手势模型；其次，引入卷积姿势机（Convolutional Pose Machine，CPM）提取交警指挥手势的关键节点，进而提取交警指挥手势中骨架的相对长度及其与重力加速度的夹角作为空间上下文特征，并引入长短时记忆网络（Long Short Term Memory，LSTM）提取交警指挥手势的时序特征；最后，设计了融合空间上下文和时序特征的交警指挥手势识别机（Chinese Traffic Police Gesture Recognizer，CTPGR），创建了包含8种交警指挥手势、时长约2小时的交警指挥手势视频库对CTPGR进行训练验证，并通过实验将CTPGR与已有交警手势识别算法进行了对比分析.实验证明CTPGR可以快速准确地识别交警指挥手势，系统对复杂背景和动态交警指挥手势具有较强的适应能力.

Visual Recognition of Chinese Traffic Police Gestures Based on Spatial Context and Temporal Features

According to the need for driver assistance systems and intelligent vehicles to quickly and accurately identify traffic police command gestures,the articulated features of traffic police gesture is firstly analyzed,and a model based on the key points and skeletons of the police gesture is established.Secondly,the convolutional posture machine (CPM) is introduced to extract the key points of the traffic police gesture.Then the relative lengths of the gesture skeletons and the angles between each skeleton w.r.t.gravity are extracted as the spatial context features of the traffic police gesture.Meanwhile,long-term memory (LSTM) is introduced to extract the temporal features of traffic police gestures.Finally,the Chinese traffic police gesture recognizer (CTPGR) based on CPM and LSTM is designed,and a two-hour traffic police gesture video is recorded to train and verify the CTPGR.Experimental results show that the CTPGR is capable of recognizing traffic police gestures with high accuracy,and is fast enough for online gesture prediction.