Multi-Level Detection Of YOLO Algorithm Based On Lightweight Network In Autonomous-Driving

Recently,autonomous driving is a hot research,and object detection is one of the basic tasks in this field.In the early stage,the object detection algorithm of autonomous driving adopted traditional detection algorithms and relied on the extraction of features manually.As the development of convolutional neural network,the deep learning detection algorithm develops rapidly,whereas there are still some problems that need to be solved in the next stage.First,deep network has many parameters,large computation,slow detection speed and large memory consumption,that making it difficult to deploy on embedded devices.Second,driving at night,the light is weak.Image resolution will be influenced,so that the vehicle detection will be difficult.Third,the accuracy of the bounding boxes' position is a little low,the vehicles need to be classified in more detail.According to these three aspects,this thesis did following researches:(1)In order to accelerate the speed of detection,in the training stage,using the lightweight network Mobile Netv2 as the backbone in the YOLO algorithm,at the same time,we adopted the dilated convolution and adjusted the weight of the small objects in the loss function to compensate for the loss of the precision caused by the lightweight network in the training stage.(2)In the detection stage,selecting the appropriate ROI area to do the two-level detection,and using the non-maximum suppression to delete the useless bounding boxes,to make up for the decreasing of accuracy caused by the lightweight network in the detection stage.Through the modifications in the training stage and the test stage,the thesis speeds up the detection without losing too much accuracy.Model gets more smaller,which provides the possibility in the embedded devices.(3)The lightweight network YOLO algorithm in this thesis is applied to the night driving detection,vehicle classification in detail and the bounding box regression.Experiments prove that,compared with traditional feature extractions manually,the method in this thesis can learn more abstract features and achieve satisfactory results at night time.The bounding box regression task realizes the regression of the bounding boxes to make the position of the bounding boxes more precise,and at the same time realizes the classification of the front car,side car and back car,building a foundation for the other tasks.