Research On Detection And Collision Warning System Of Pedestrian And Cyclist Based On Deep Learning

Cars are becoming the most common means of transportation.It's convenient for people's travel,but it also causes frequent traffic accidents.As vulnerable groups of traffic environment,the security problem of pedestrian and cyclist can't be underestimated.Therefore,much focus has been placed on establishing a perfect detection and warning system of pedestrian and cyclist.Besides,deep neural network has been widely applied to machine vision due to its excellent performance.The application of deep learning to driver assistance system is becoming the development trend.In this paper,pedestrians and bicycles are research targets and the front collision warning system is designed based on deep learning.The system gets environmental information through the video camera and extracts the vehicle speed from GPS module.Then an early warning model is built to judge whether the target is safe.The main work of this paper is as follows:(1)A target detection model based on deep learning is built.YOLOv2 network is used as the basic model of target detection.In order to improve the detection accuracy of cluster small targets,ResNet is added to YOLOv2 to form a new model called YOLO-R.Then the sample database of pedestrian and cyclist is constructed.The size of anchor boxes and other network parameters are modified.After this,the trained model is more suitable for pedestrian and cyclist.The matching algorithm is used to classify pedestrian and cyclist further,and Kalman filter is utilized to achieve multi-target tracking.(2)An algorithm for detecting the relative distance between target and test vehicle is developed.The monocular ranging algorithm based on inverse perspective transformation and data regression modeling is used to measure the distance of the leading targets.Firstly,some pixels of original image are transformed into the world coordinate system through inverse perspective transformation and IPM image is obtained.To avoid a large number of repeated computation process of the inverse perspective transformation and increase the range of distance measurement,the regression model of the pixel coordinates of original image and IPM image is set up,and the distance is estimated through the linear relationship between the pixel coordinate of IPM image and the world coordinate.Then,the pitch angle of camera is calculated in real time by the road vanishing point detection algorithm based on the texture orientation estimation,and pixel coordinates of targets are amended to overcome the effect of pitch angle's change in distance measurement.(3)The fuzzy warning system is established.The warning activation area is set to exclude some securities.Then fuzzy synthetic evaluation method is applied for target's warning in the warning activation area.According to the position of targets,the lateral and longitudinal distances,the vehicle speed and time to collision,the system judges the degree of danger and presents the warning level in the video image.(4)Target detection,distance measurement and early warning functions are validated.First of all,target detection experiments are conducted and YOLO-R is compared with YOLOv2 in detection results.Experimental results show that YOLO-R achieves higher accuracy and recall rate,its mAP is increased by 3.4%.Then,in static and dynamic ranging experiments,the measurement error of the lateral and longitudinal distances is less than 7%.Finally,the early warning experiment is completed.The results of warning systems with and without early warning activation area are compared,and the reasons for false and missed alarms are analyzed.The experimental results show that the accuracy and real-time of this system can meet the requirements of the front collision warning system.