Study On Load-related Geometric Feature Recognition Of Heavy Vehicle Based On Deep Learning

China is the country with the largest number of bridges in the world,but many bridges have problems of long service life and poor safety condition,which are in urgent need of maintenance and management.In the field of bridge maintenance,the bridge health monitoring system is a more efficient method than the manual overload checkpoint.By monitoring the response information of the bridge structure caused by heavy vehicles in real time,the situation of overloaded vehicles crossing the bridge can be found in time.However,the existing bridge health monitoring system cannot determine which heavy vehicle caused the vibration monitoring data of the bridge structure,nor can it obtain fine vehicle information such as speed and wheelbase.In response to these problems,this thesis used a target detection algorithm based on deep learning to recognize the load-related geometric feature information of heavy vehicles crossing the bridge,assisting the existing bridge health monitoring system to obtain more accurate structural response information.The main work and innovations are as follows:(1)A heavy vehicle dataset was produced.By collecting monitoring images of real roads in Hangzhou as material,a heavy vehicle dataset containing different weather and different time periods was produced,with a total of 15,000 images.(2)An improved RC-ZF algorithm based on Faster R-CNN was proposed.By analyzing the difficulties of target detection on real road monitoring images,the Faster R-CNN algorithm was improved to promote the detection ability of wheel targets.The improved RC-ZF algorithm was verified on the heavy vehicle dataset.The results show that the mAP(mean average precision)of RC-ZF algorithm can reach 90.84%in daytime and rainy weather,which is about 3.4%higher than the original ZF network,and the model recognition speed reaches 0.051s each frame,basically meeting the requirements of real-time detection;at night or in the early morning,the mAP can reach 85.54%and the detection speed remains unchanged.(3)A fine vehicle recognition algorithm based on plane ranging was proposed.After the RC-ZF algorithm recognized the characteristic information of the heavy vehicles,a simplified camera calibration method was used to complete the plane ranging to further recognize the fine vehicles.The real information of vehicles and the output information of the algorithm were compared through experiments,and the results show that the recognition accuracy of this method can reach 96%and 97%for the axle number and lane information of heavy vehicles respectively;for the wheelbase information,the error rate can be controlled within 7%;for the speed information,the error rate is within 5%.The algorithm has high recognition accuracy.(4)A vehicle recognition display system was designed and implemented.Users can log in to the system through the Web interface to upload the image to be detected,and the system displays the final recognition result through the Web interface.