Research On Monocular Ranging Method Based On Deep Learning

With the development of social economy and the progress of science and technology,national car ownership also increase,although has brought a lot of convenience to people’s life,but also associated with increases in congested roads and serious accidents,the car driving safety has become the current auto industry development is facing a prominent problems and challenges,Distance measurement of road targets is an important part of driving safety.In recent years,with the rapid development of deep learning,road target range has become a hot research topic in the field of automotive safety driving,so this article will target detection and monocular visual range combined with this paper proposes a new single visual distance algorithm based on deep learning target detection,Furthermore,the target detection algorithm based on deep learning is improved and compressed to adapt to edge computing devices with less computing power.The main research work and innovations are as follows:(1)Way to target range,this paper proposes a new single visual distance algorithm based on deep learning target detection,target detection algorithm YOLOv5 x tests first goal on the road,road target is obtained on the test frame of reference point of pixel coordinates,then zhang camera calibration algorithm to establish the mapping relationship between the world coordinate system and pixel coordinate system,By constraining one dimension in the world coordinate system,the pixel coordinates of camera and road target are converted into world coordinates,and then the real horizontal distance between them is calculated.(2)Due to the accuracy of target detection is one of the important factors affect the distance measuring result,so this article to improve the YOLOv5 x network,the introduction of coordinate mechanism of attention for its network of feature extraction was improved,from two different directions to global pooling of network channel,in global compression features at the same time,also extract the location information,can effectively extract the feature information,The detection accuracy is further improved,and then the improved YOLOv5x-CA network model is compared with the basic network through experiments.The experimental results show that the m AP of the improved network model is improved by 2.1%.(3)Aiming at the problem that the improved YOLOv5x-CA network model has high computational power requirements and is difficult to deploy on low performance platform,this paper uses an adaptive threshold pruning algorithm to perform network pruning on YOLOv5x-CA model to reduce the number of model parameters and model size,and takes the scaling factor learned after sparse training as the measurement parameter.According to the change of scaling factor distribution,the pruning threshold was adaptively determined to obtain the lightweight model and reduce the number of model parameters and model size.The experimental results show that the number of parameters in the compressed YOLOv5xCA-Pruning model decreases by 88.2% and FPS increases by 66.5%.Finally,YOLOv5xCA-Pruning is used to detect the targets on the road and distance measurement is carried out for the road targets.Experimental results show that the distance measurement error rate of the proposed single-eye distance detection algorithm based on deep learning target detection is within 5.7% for the road targets within 30 meters.