Railway Hand Signal Gesture Recognition Based On Depth Data Features

Railway shunting operation,as an important section in the process of railway transportation,is a highly technical and complex task.It is a significant symbol to measure various indexes in the transportation.While training shunting operators,to accurately identify and display hand signals is of great importance.So far,hand signal training has still relied on traditional artificial teaching model.However,because of a large number of training members,as well as various kinds of hand signal types,the training efficiency is relatively low.A gesture recognition method based on artificial intelligence is applied to railway shunting in this thesis.The gesture feature datas are collected through residual neural network,and the color of signal flags is combined with CamShift algorithm for dynamic tracking and recognition.The integrated feature datas are sent to the classifier for autonomous training and recognition of hand signals.The primary four contributions of this thesis are addressed below.(1)The hand signal sample datas are collected.Through careful observation of hand signals,and combination of the gesture feature of railway hand signals,the railway hand signals are divided into two categories,namely static and dynamic,among which the static category include ten types of static pictures,and the dynamic category include ten types of dynamic videos.(2)The human gesture joint points recognition method is proposed.In this thesis,open source deep learning network model is applied to extract joint points,and residual neural network is used to find out all the candidate joint points.Relevance between joint points is taken as the weight between points.Graph theoretic node clustering method is adopted to determine to whom the joint points belong to.Finally,Integer Linear Programming(ILP)is used to work out the model.(3)The color recognition and tracking algorithms of the signal flags are designed.The signal flag colors of railway hand signal are divided into three,namely red,green,and yellow.This design uses CamShift algorithm to identify the color of signal flag,and dynamically track the action of signal flags.By constantly calculating the size and center position of search window,and iteratively calculating the initial value,the difficulty of target deformation and occlusion can be effectively solved.(4)The hand signal gesture classification and analysis model is established.The pre-processed datas are sent into K-Nearest Neighbor(KNN),Logic Regression,Support Vector Machine,and other classifiers.The model is optimized by continuously searching parameter through cross-validation and grid search.Finally,different algorithms are compared and analyzed.The results show that the performance of Support Vector Machine is better in static and dynamic category,and the accuracy can reach 98.999%and 97.999%respectively.