Research On Vehicle Detection Based On Convolution Neural Network And Late Fusion Algorithm

Vehicle detection is an important research direction in the field of autonomous driving and intelligent transportation system.Vehicle detection not only plays a role in reducing accidents in the field of autonomous driving,but also helps solve the problem of traffic pressure in the field of intelligent transportation.Vehicle detection late fusion technology,based on the data of vehicle detection,has also become one of the hot directions in the field of object detection because it can increase the mAP of vehicle detection.To solve the low accuracy of vehicle detection in China,this paper proposes a vehicle detection technology on the basis of convolutional neural network and late fusion algorithm,aiming to improve the mAP of vehicle detector.The main contents of this paper are as follows:1)Preprocess dataset and establish evaluation method.KITTIplus is created to address the problem of lacking of certain vehicle classes from the KITTI dataset.To solve unbalanced distribution of classes in KITTIplus,data augmentation is proposed to increase the amount of the vehicle class which is far less than other c lasses before augmentation.On top of this,an objective evaluation method is designed to calculate the mAP of the detector,and a subjective evaluation method is designed to validate the practicability of the vehicle detector in the domestic scene.2)Design SSD-V vehicle detector based on Single Shot MultiBox Detector.The designs of SSD-V mainly includes the network,default boxes and loss function.The vehicle detector adjusts the hyperparameters while it is training in KITTIplus,to obtain relatively optimal network weights.This study will calculate mAP of SSD-V on objective evaluation test and save the test data.Furthermore,we validate the model on subjective evaluation test to verify its practicability in domestic scenarios.3)Design YOLO-V vehicle detector on the basis of YOLOv3.The designs of YOLO-V mainly includes the base network,prior boxes and multi-scaled training method.Besides,this study will train the model by setting multiple sets of hyperparameters on KITTIplus to gain the relatively best model weights and calculate the mAP of YOLO-V.Also,this study will test YOLO-V on subjective evaluation method to show the practicability of model.4)Proposed the ALFA-V based on the design ideas of multiple late fusion algorithms.The design of ALFA-V mainly includes clustering method,class of the cluster and object localization.Firsty,I input the detection results of SSD-V and YOLO-V into ALFA-V and get the relatively best mAP of ALFA-V on KITTIplus by adjusting multiple sets of hyperparameters.Then,this study will compare the mAP of SSD-V,YOLO-V,NMS and DBF to demonstrate the superiority of ALFA-V in vehicle detection.The research shows that the basic detectors SSD-V and YOLO-V in this paper achieved 76.68% and 83.48% mAP respectively in KITTIplus.This paper verifies that the late fusion algorithm of vehicle detection based on convolutional neural network can increase the mAP of detection.ALFA-V finally can improve the mAP to 84.51% in vehicle detection,which has certain advantages compared with other late fusion algorithms.