Research On Pedestrian Vehicle And Driver Posture Detection System Based On Machine Learning

With the increasing number of car ownership and the improvement of urban road construction,road safety problem has become an important factor of threatening personal safety.In this context,the Advanced Driver Assistance System has come into being.As an important component of the advanced auxiliary driving function,detection system of pedestrian,vehicles and drivers poses is mainly used to feedback the traffic information to the drivers in real time and supervise the driver's driving state.As to the detection of pedestrian and vehicles,participants of road traffic are numerous,and the road scene is complex,so it is a difficulty to quickly and accurately identify the pedestrians and vehicles on the road to help drivers drive safely.As to the detection of drivers poses,it mainly concerns for the head posture monitoring which is the most important factor for safety driving.Although it is lower demand of algorithm accuracy of this scene to pose monitoring,it is still need to go further study about how to quickly get the head pose information,and warn the driver as soon as possible before the accident.As the road traffic scene is special,the general object detection algorithm cannot meet its detection accuracy and time consuming requirements.In the head pose monitoring,the existing methods often have a higher accuracy of pose detection,but the real-time performance is difficult to meet the actual needs.In the light of these problems in the auxiliary driving system,this paper carries out the following works:(1)propose a method of Faster r-cnn pedestrian vehicle monitoring with no loss of characteristics.For features loss problem of Faster r-cnn network in pedestrian vehicle detection,it could output the fixed dimension feature coding by using the random forest classifier,which would replace the spatial cone pooling layer in the original Faster rcnn network.Experimental results show that compared with the original Faster r-cnn network and the classification model which used the random forest classifier to replace fully connected layer of Faster r-cnn,the accuracy of pedestrian vehicle detection has been improved at a certain extent.(2)propose a method of detailed information enhancement for pedestrian detection network compression.In view of the "knowledge formulation" compression method ignoring the details of information,this paper will quantify the sample apparent information,and join into the network "learning" goal.Experimental results show that this method can effectively compress the pedestrian detection network,and its compression accuracy is better than the traditional "knowledge extraction" algorithm.(3)propose a method of fast estimation of the driver's head posture,which is based on dot pitch of facial features.Aiming at solving the problem of poor real-time estimation of the existing head pose,based on advantages of accurate modeling method for detection of facial feature points and the fast estimation method of facial feature,this paper proposes fast head's pose estimation based on the dot pitches of driver's facial features.The experimental results show that this method has better real-time performance compared with the existing method,when it meets the accuracy requirements of the driver head pose in monitoring scene.