Research On Object Detection And Measurement Based On Monocular Vision

With the continuous growth of vehicle ownership,our country is facing an increasingly severe traffic safety situation.The development of Advanced Driving Assistance System(ADAS),which can perceive and analyze the environment around the vehicle and provide early warning of potential hazards,is of great significance to ensure driving safety.Real-time and accurate acquisition of the position and size of objects such as pedestrians and vehicles on the road is a prerequisite for ADAS to make decisions.The measurement method based on monocular vision has attracted wide attention in recent years because of its advantages of rich information acquisition and high real-time performance.Since object detection is the basis of object measurement,the accuracy of the detection algorithm seriously affects the accuracy of object measurement.This paper makes full use of the advantages of deep learning and proposes a high-performance object detection algorithm in driving scenes.Combined with the object detection results,this paper proposes a monocular vision object measurement algorithm based on geometric relationship derivation.The main contents are as follows:This paper improves the YOLOv4 object detection algorithm based on the statistical characteristics of the KITTI traffic scene dataset.Aiming at the problem of large image aspect ratio and uneven sample distribution in this dataset,a Ladder data augmentation strategy is proposed,which avoids the problem of image distortion and loss of a large number of samples caused by Mosaic enhancement.Aiming at the problem of the imbalance in the number of samples of different categories in this dataset,a large-category undersampling strategy is used to load images during training.In addition,the introduction of the Single Stage Headless(SSH)context module further increases the feature extraction capabilities of the model.The experimental results show that the detection accuracy of the improved YOLOv4 algorithm on the KITTI dataset is 46.4% higher than the original algorithm,and the detection accuracy of different categories is more balanced.Based on the existing geometric ranging model,this paper proposes a monocular measurement algorithm based on geometric relationship derivation.The algorithm selects reference points based on the output bounding box of the improved YOLOv4 algorithm,and can simultaneously measure the distance and size of the target object.Since the internal and external parameters of the camera have a greater impact on the measurement accuracy,the camera needs to be calibrated before measurement.Experimental results show that the measurement algorithm proposed in this paper has high measurement accuracy and speed,is suitable for object measurement in vehicle driving scenes,and is of great significance for the development of high-performance ADAS.