Research On The Vehicle Mounted Collision Avoidance Warning System Based On Asphalt Pavement Condition Detection

Forward Collision Warning(FCW)has been shown to be effective in reducing the occurrence of rear-end vehicle crashes.However,in unfavorable asphalt pavement conditions,such as wet,snow and ice,the braking performance of vehicles is reduced,increasing the likelihood of rear-end collisions.To improve the warning effectiveness of FCW under different asphalt pavement conditions,this thesis proposes an in-vehicle collision avoidance warning system based on real-time asphalt pavement condition detection.The system uses image analysis methods to realize dynamic multi-stage collision avoidance warning based on the type of pavement condition to improve warning performance under different asphalt pavement conditions.Considering the difference of pavement adhesion coefficient under different asphalt pavement conditions,the asphalt pavement conditions are classified into four types: dry,wet,snow and icy.The images of different asphalt pavement conditions are collected,and the asphalt pavement condition image dataset is established after image enhancement,region of interest extraction and data augmentation processing.The Mobile Net v3 model is used as the base model,and an improved efficient channel attention module is designed for low-cost channel weight assignment.Additionally,dilated convolutions are introduced at the front end of the model to increase its overall perceptual field.By combining these techniques,a Mobile Net(DH-Mobile Net)asphalt pavement condition recognition model is established to achieve the effective recognition of the above four asphalt pavement condition images.The Labelme image annotation tool is employed to annotate the tail area of vehicles and establish a vehicle target image dataset.The YOLOv5-nano model is chosen as the base model,and the Si LU activation function is replaced with the hard fitting function h-swish to reduce computational costs.To enhance target detection accuracy,the SIOU bounding box regression loss is introduced in place of CIOU loss.A YOLO(HS-YOLO)vehicle target detection model combining the hard fitting function and the optimized bounding box regression loss is established to realize the effective detection of the front vehicle target.The geometric relationship conversion method is used to realize front vehicle ranging based on the lateral width of the front vehicle,and the ranging accuracy is verified.In addition,the effective warning area is delineated using the display area of the lane ahead in the image.The midpoint of the lower edge of the vehicle detection frame is used as the reference point of the vehicle position,and the effective warning area is used to determine the effective warning target,thereby improving the warning accuracy.A dynamic multi-level collision avoidance warning model is proposed in this thesis,which considers both the type of asphalt road state and the state of the front vehicle.To achieve this,the self-vehicle speed and the relative vehicle speed are used to calculate the front vehicle speed,which is then used to determine the front vehicle’s motion state,including stationary,uniform speed,and braking.Different safety distance models are established for each of these states.Furthermore,the pavement condition-pavement adhesion coefficient mapping model is developed based on actual measurements and literature review.The model selects the corresponding pavement adhesion coefficient and the minimum stopping safety distance based on the type of asphalt pavement condition.Different safety warning strategies are adopted based on the detection results of the front vehicle state and asphalt pavement condition to realize dynamic multi-level warning.Mathematical analysis and joint simulation of Car Sim and Simulink have demonstrated the effectiveness of the collision warning model combined with the asphalt pavement condition in unfavorable road conditions.The system is developed and validated on the Jetson platform for both hardware and software.The DH-Mobile Net asphalt pavement condition recognition model and the HSYOLO vehicle target detection model are deployed with Tensor RT acceleration on the hardware platform to improve the real-time performance of the system.An experimental platform for a vehicle collision avoidance and warning system based on asphalt pavement condition detection has been established.The effectiveness of the system has been proven by verifying distance measurement accuracy and warning function through both static and dynamic vehicle experiments.