| As the development focus of smart cities and smart transportation,autonomous driving technology is getting more and more attention.Our current autonomous driving technology is in the transition period from L0 to L5,which requires a process when transitioning to the highest level,completely controlled by the system.In this transitional period,the command and control of traffic police comrades and drivers must be inevitable.Aiming at the problems of pedestrians’ interference with traffic police detection in the application environment,the distance of shooting distance and the camera’s jitter are not recognized,and the gesture track dimension is high,the data volume is large,and the real-time performance is poor.This paper designs the following schemes:First,identify and locate the traffic police.The identification and location of the traffic police as a precondition for traffic police gesture recognition,the accuracy of its identification is directly related to the practicality of the project.This project,combined with the actual application scenario,decided to use the clothes and legs of the traffic police as the detection target.Use the Haar-Adaboost method as the front part of the cascade classifier.The improved Sobel feature is then utilized,and HSV color features and contour features are used to cascade the latter portion of the classifier to train the resulting cascade classifier.The final experimental results show that compared with the traditional single Haar-Adaboost detection method,the average detection rate of this method is increased by 8.1%,and the average false detection rate is the most obvious,from the original 41.5% to 9.2%.After obtaining the location information of the traffic police,the project has designed a low computational complexity and high real-time performance for the problem that the shooting distance is far away in the actual environment and the resolution of the image caused by camera shake is too low and unclear.Image super-resolution algorithm: First,image registration is performed to obtain sub-pixel displacement information between adjacent two frames of images,and then a higher resolution image is obtained using weighted nearest neighbor interpolation.Finally,the image is deblurred by Wiener filtering to obtain a subsequent high resolution,high resolution image.Lay the foundation for later hand tracking and recognition.The hand detection part adopts a deep learning method: training the CPM network to obtain a network model for detecting the position information of the human key points(both hands and head),and then calculating the relative positions of the hands and the head.Good results are difficult to obtain due to the simple training of the model to detect the hand.Generally,due to the interference of light and skin objects,the recognition accuracy is low and the false detection rate is too high.Therefore,the second-order form of Kalman filter(?-? filter tracking)is used to track and filter the gestures in the continuous video stream.Improve the accuracy of hand detection and lay a foundation for more accurate gesture track recognition.The final experimental results show that the key point detection rate has increased from 91.4% to 97.9%.Finally,in the gesture trajectory recognition portion,the hidden Markov model is trained by using the relative displacement information of the obtained hand relative to the head as a feature.In the experimental process,we start with a simple problem and gradually deepen the way.Firstly,the experiment is carried out with one hand.At the same time,the complex high-dimensional gesture sequence information is simplified by quantifying the direction and displacement of the gesture track,which improves the recognition accuracy while reducing the feature dimension.After the one-hand gesture model is obtained,the gesture trajectory recognition experiment of both hands is performed.The experimental results show that the recognition rate of the eight traffic police commanding gestures reached 98.0%. |
PDF Full Text Request