Research On Intelligent Vehicle's Motion Perception And Street Intention Prediction For Pedestrians

With the advancement of science and technology and social development,intelligent driving technology came into being.Environmental awareness is the fundamental premise of intelligent driving,providing the vehicle with the right visual information.As an important participant in the traffic environment,pedestrians are flexible and changeable,which can easily cause unnecessary traffic accidents.The pedestrian's state of motion(position,speed and acceleration)affects the intention of crossing the street,combined with the movement state of pedestrians and other influencing factors.In the case,you can predict its crossing intentions to avoid unnecessary accidents.Therefore,this paper mainly studies from the aspects of pedestrian and its head direction detection,pedestrian tracking and motion state estimation and pedestrian crossing intention prediction.1.Pedestrian and its head direction detection.Based on the deep learning platform Caffe,the Faster-Rcnn and Fast-Rcnn network frameworks are built to collect a large number of scene images containing road pedestrians.Labelling is used to label the scene images,and a sample set is created.The network is trained to detect pedestrians and A network model of its head orientation.Comparing the detection effects of different feature extraction networks with different network frames,the Faster-Rcnn+VGG16 combination has the best detection effect,and the overall accuracy is close to 90%.2.Pedestrian tracking and its motion state estimation.The color distribution model is established,and the color distribution is transformed into particle filtering.Combined with the edge-based moving image features,the prior probability of particle filter calculation is updated by the Pap address,and the pedestrian effect is tracked and the tracking effect is good.The model establishes the pedestrian pedestrian motion state model and measurement model of the vehicle camera,determines the system matrix of the motion state,the noise matrix and the state transition matrix,and uses the particle filter tracking model to achieve the purpose of motion state estimation,realizing the pedestrian longitudinal speed and the person.A good estimate of the horizontal and vertical distance between the cars.3.Pedestrian crossing intentions are predicted.Under the condition of comprehensive consideration of various pedestrian crossing factors,the node variable parameters are finally determined and a pedestrian crossing intent prediction model based on Bayesian network is established.The structure learning is used to determine the relationship between each node variable,and the data is correlated and analyzed by the statistical product and service solution software(Batstatical Product and Service Solution)and the function Tab.ular_CPD in the Bayesian Network Toolbox(BNT).The prior probability is initialized,and the conditional probability table of each node variable obtained by parameter learning is compared with the probability obtained by the sample data.Finally,the joint tree inference algorithm is used to obtain the influence of each node variable on pedestrian crossing,after given data.Using the obtained model to reason,the pedestrian crossing probability is obtained,and combined with the actual experience,it can be known that the probability of crossing the street prediction is reasonable and meets the actual requirements.4.Monocular ranging verification experiment.The purpose of the experiment is to verify the accuracy of the distance between the pedestrian and the front of the vehicle at different installation heights and to detect the accuracy of the rectangular frame obtained by the pedestrian.The pedestrians collected by the experimental platform are detected,the segmentation effect diagram of the detected pedestrian rectangle frame is extracted by color segmentation,and the pixel coordinates of the midpoint of the bottom edge of the rectangular frame are obtained,and the pedestrian is obtained according to the small hole imaging model and the proposed solution.The distance from the front of the car,in the experiment,the height of the camera is installed to a different value,and the picture captured by the on-board camera is detected,and the calculated distance value obtained is compared with the actually measured distance value to determine the installation height.And detecting the error caused by the model on the measured distance.The distance error caused by the pixel error is relatively small and the height is the main error of the ranging,and the error of the camera when the installation height in the vehicle is 1.36m is relatively small.