Research On Collision Warning System For Trucks Based On Machine Vision

In recent years,the rapid development of road transportation industry,trucks in the transport industry has assumed an extremely important task,but also bring more serious transport safety problems.Factors such as fatigue driving,over-speeding and over-loading,and restricted rear view mirror vision make it very easy for rear-end and side collision accidents to occur when trucks are driving.In view of the above problems,this topic combines the demand of improving truck transportation safety,and designs and develops a truck collision warning system based on machine vision.The main research contents are as follows:(1)Image enhancement method research and improvement.The early warning system is based on monocular vision to complete target detection,distance measurement and early warning.Firstly,to address the problem of poor image quality and low contrast of CMOS sensor-based monocular camera in poor and low illumination environment,the principle of dark channel a priori enhancement algorithm is analyzed and the image partitioning process is realized by improving the adaptive correction of transmittance to reduce distortion and Halo phenomenon,and the histogram equalization algorithm is used to denoise and enhance the contrast of the small dynamic range area.The improved algorithm is suitable for image enhancement in different illumination environments,and obtains a larger amount of image information and retains more image feature details;the timeliness is reduced by 35% compared with the original dark channel a priori algorithm in terms of average frame processing time.(2)Construct and optimize the SSD-based target detection model.Based on the completion of image pre-processing,the SSD network structure is analyzed and improved to achieve efficient detection of vehicle and pedestrian targets.The target detection model is first built based on the lightweight feature extraction network Mobile-Netv2 replacing the SSD backbone network VGG16,then the feature enhancement module is fused and the a priori frame sieving mechanism is improved to optimize the detection capability of the model and reduce the computational scale,and finally the dataset is built and the network parameters are configured based on KITTI and PASCAL VOC.The improved algorithm reduces the computation size and increases the number of frames detected.In the KITTI test set,the m AP value increases by 2.13% and 0.5% compared with the SSD model and Mobile-Net+SSD model,respectively,demonstrating the accuracy and generalization ability of the improved model in multi-environment scene detection.(3)Monocular visual imaging and ranging model building.Based on the completion of target detection,monocular ranging models such as inverse perspective transformation were analyzed and compared,and the monocular ranging model was selected based on camera calibration method considering the stability and accuracy of ranging.Firstly,the monocular vision imaging model is analyzed,then the monocular ranging model is established based on target detection classification by calculating the internal and external parameters of the camera and correcting the image distortion based on the calibration plate.Finally,we analyzed and evaluated the ranging effect of the model based on KITTI dataset,Open CV plus fog image dataset,and field ranging experiments,and verified the accuracy and stability of the ranging model using the coefficient of determination,root mean square error and other indexes.(4)Fusing image pre-processing,target detection ranging model and warning strategy to build a collision warning system.Based on the Trucksim vehicle dynamics software,the internal wheel difference data of F590 truck is calculated and simulated to verify,and then the TTC relative collision time and safe braking distance warning models are analyzed and the fusion warning strategy is proposed to divide the forward and lateral warning areas of the vehicle for the problem of easy false alarm by a single warning strategy.Based on the above research,a monocular vision-based collision warning system for trucks was designed,covering forward and lateral collision warning,and the collision warning system software was designed using pyqt5.After experimental verification,the warning system obtained 90% and 83.3% warning accuracy in the lateral inward wheel difference blind zone and forward and right turn collision warning experiments respectively,and it can achieve efficient and stable warning in low illumination environment.The collision warning system designed in this study can improve the safety of truck driving,effectively reduce the occurrence rate of forward and lateral traffic accidents of trucks,and provide guarantee for transportation safety.