Light-weight Mobile Road Damage Detection Algorithm Research And Implementation

Road is the city infrastructure,rapid economic development,social progress of the fundamental guarantee,timely and effective road damage detection work can not only extend the life of the road,but also to protect the safety of vehicles,to maintain the safety of people’s lives and property,with significant economic value and social benefits.However,with the increasing maintenance mileage and traffic flow,the task of road surface damage detection is increasingly arduous,and there is an urgent need for an intelligent,lightweight road damage detection method.In order to improve the automation of road damage detection,researchers began to study deep learning road damage detection methods,which reduce the cost of human labor,improve the efficiency of road maintenance in a certain sense,and provide theoretical support for the digital construction of road maintenance,however,some unresolved issues remain:(1)Too much bias towards theoretical class research,the data set and experimental part are too idealized,not taking into account the influence of various factors that may occur in the actual road damage detection;(2)The hardware equipment of the experimental environment is generally high,and the maintenance department may not be able to experiment with the required hardware conditions in the actual application process,and the model faces certain difficulties in landing.To address the above issues,this paper develops a technical study combining deep learning target detection and model light-weight methods for computer vision,while the relevant technology for hardware and software system integration,the development of lightweight mobile road damage detection prototype system,to provide the basis for road maintenance,to promote the development of high quality road maintenance.The main research contents of this paper are as follows:(1)For the road damage type characteristics and the actual detection process has been the problem of water damage,oil and other interference,in terms of network model design,this paper builds a damage detection network resistant to non-target object interference.The network model uses Ghost Conv module in the coding network to linearly amplify the extracted features to enrich the shallow feature information of road damage while reducing the computational effort;the feature fusion mechanism with multiple attention constraints is added in the decoding network to achieve the focus on the damage information in the feature map through the method of multi-layer feature attention constraints and fusion.Experiments show that the network model achieves 91.6% m AP and is able to achieve accurate detection of damage under the interference of water stains,oil and shade.(2)To address the problem of accuracy decay of the deep learning network model in the process of lightweighting,this paper proposes a road damage model lightweighting method based on dynamic weight pruning,which retains branches with low weights but important information,and integrates pruning and fine-tuning to improve the accuracy and time efficiency of compression.The experiments show that the proposed road damage lightweighting method compresses the number of parameters by 43.9% while losing only 4.6% of accuracy,which can achieve significant lightweighting of the network model while preserving accuracy.(3)Combined with theoretical and technical research content,the integration of hardware and software,the development of lightweight mobile road damage prototype system,to achieve automatic detection of damage,while the damage information to locate,generate detection trajectory route,and automatically export the detection report at the end of the detection,to provide reference for the practical application of lightweight mobile road damage detection algorithm,a single frame detection speed of up to 23.3 ms to meet the demand of real-time detection.