Research On Object Detection Method Of Traffic Scene Based On Lightweight Neural Network

Machine vision-based object detection is one of the hottest research directions in the field of autonomous driving.In recent years,many excellent deep learning-based object detection algorithms have continuously emerged.Among them,YOLOv3 object detection algorithm is an excellent end-toend network in terms of detection accuracy and speed.The YOLOv3 algorithm can perform object detection on pictures of any resolution,but the bounding box position regression calculations in the network training process and the prediction stage have accuracy deviations,as well as the problem that the model occupies huge running memory and storage space.Therefore,designing a lightweight YOLOv3 network to accurately detect vehicle targets in complex traffic scenarios is in line with actual engineering needs.The followings are the specific methods for the research and improvement of the original YOLOv3 network in this thesis:(1)For YOLOv3 network,if the rectangular image in the traffic scene has been stretched,the image deformation would reduce the accuracy of detection.In this thesis,the anchor frame size is obtained by clustering with the use of K-Means algorithm,and the size of the pre-selection box mapped on each layer of prediction feature maps is set by the calculation formula.This method effectively improves the network’s low accuracy in detecting traffic scenes with rectangular views,and reduces the time consumption of redundant operations in the model training process of candidate box parameters.(2)The number of preselected box category labels generated for images of traffic scenes is uneven,which easily leads to false detections and missed detections.The adopted method uses the focus loss function to solve the problem of sample label imbalance.This improves the detection accuracy of the model on the Pascal VOC dataset.The detection effect of mean average precision of the improved YOLOv3 network model is 2% higher than the original YOLOv3 network.(3)To improve the part of the feature extraction network with the most traditional convolution operations in the YOLOv3 network,a lightweight feature extraction network design method was adopted.This method is improved in combination with a lightweight feature extraction network while maintaining the resolution of the input picture.The size of the trained model is only one-ninth the size of the original model,and the reduced accuracy of the model is compensated by the improved candidate box overlap computer system.This thesis names this design network model FM-YOLOv3.Finally,the detection experiments of the KITTI data set of traffic scenes are compared.The detection results of the improved FM-YOLOv3 network model show that the lightweight design network model research method proposed in this thesis has reached the expected effect level in model memory size and detection accuracy.