| With increasing traffic accidents,autonomous vehicles have attracted more and more attentions in recent years.Intelligent object detection can effectively find surrounding obstacles and determine their locations,which is a key technology in intelligent driving.Detection based on radar and other sensors may lead to much high cost and the over sensitivity problem in practice.In this thesis,a monocular on-road object detection system is established.The main work includes: construction of the hardware platform,object detection in images,detection model compression,object distance estimation and data transmission.The details are as follows:(1)Hardware platform construction of monocular on-road object detection.Based on the detection requirements,a Point-gray camera is used for image acquisition.The camera is installed on top of an intelligent vehicle and camera calibration is performed.The main algorithms are executed on an industrial computer with Python language.(2)Object detection in two-dimensional images.A total of 53 categories of common obstacles in road scenarios are detected based on YOLOv3 algorithm,such as motor vehicles,pedestrians,traffic lights,traffic signs etc.Different deep learning frameworks are compared in the experiments,and finally Darknet is adopted as the detection framework.Based on the hardware platform requirements,a channel pruning method is carried out to prune the detection model after sparse training,and the new model is finetuned,which has reduced the size of the model to 1/10,and the detection accuracy has not decline significantly.(3)Object distance estimation.Based on the two-dimensional detection results in the images,a distance estimation method is developed to estimate the three-dimensional positions of each object.With object segmentation,the distance estimation accuracy has been improved.With the third-order exponential data smoothing algorithm,the stability of the estimation results due to the jitter of sample point in a video sequence has been improved.In addition,data transmission is implemented to send the detection results to other modules.UDP protocol has been used in data communication to ensure the real-time and reliability of data communication.In this thesis,an on-road object detection system based on a monocular camera has been established,and object detection algorithms has been developed for application of the system,which lays a foundation for further research on sensing and recognition in intelligent vehicles. |
