Inlligent Analysis Of Traffic Events Based On Software-defined Cameras

With the hyper growth of China’s road traffic network in recent years,the importance of road safety is becoming increasingly apparent.Intelligent analysis of traffic events has become the trend in solving the current road safety problems.Weather classification and vehicle detection,which is the cornerstone of traffic incident analysis,play a very important role in intelligent surveillance.In addition,current intelligent analysis of traffic events is mostly based on high-performance GPU servers for computing whose deployment costs are relatively high.In view of the above situation,this paper first implemented the classification algorithm of road weather and the detection algorithm of vehicles,then transplanted the algorithm to the relatively inexpensive software-defined camera,finally used the embedded chip to do the intelligent analysis work of road weather classification and vehicle detection.Aiming at the task of road weather classification,this paper constructed multi-class road weather data set which has adverse weather conditions,then selected MobileNetV3 as the classification algorithm from the state-of-the-art algorithm,then improves the characteristic learning ability of the model by means of knowledge distillation and spatial attention mechanism,and finally evaluated the model and do the ablation study on our dataset.The result of the experiments verified the effectiveness of the proposed method,and obtained the transplanted model can be actually put into use.In the road vehicle detection task,the road vehicle data set was constructed from the actual road monitoring videos.This paper reviewed the state-of-the-art method in objection detection area and selected the YoLov3 algorithm as the road vehicle detection method according to the actual task requirements and hardware equipment limitation.A new structure of object detection model for transplantation is proposed based on YoLov3.First a new detection layer using large-scale feature maps is added to improve the model capable of detecting small targets,then bypass connection layer and dense connection layer is added to improve feature utilization and the detection accuracy of our model.The proposed model was converted into a model file that can be used by the softwaredefined camera.Depending on the software and hardware characteristics of the softwaredefined camera,the weather classification algorithm and target detection algorithm for the software-defined camera were developed respectively,and the model was successfully transplanted and used.The necessity,effectiveness and practicability of the model improvement were verified through the experiments.