Research On Real-Time Detection Of Video Object Based On Improved YOLO Model

Thanks to the rapid development of computing equipment,information technology and artificial intelligence,unmanned driving has attracted extensive attention from the industry and academia.It integrates the latest research results in various fields such as environmental perception,positioning navigation,and planning and control.The biggest difference between the unmanned system that began after the 2004 DARPA Challenge and the last century's unmanned system is that today's unmanned system is not dependent on supporting facilities,and the focus of research is more on how to make vehicles understand the world independently and control the vehicle based on this..It can be said that the extent to which the driverless car senses and understands the world determines the advanced level of the driverless system.The environment perception is mainly responsible for the acquisition of many information and calculating the semantic information of the objects around the vehicle,including the distance of the front vehicle,the indication of the traffic light ahead,the number on the speed limit sign,the curvature of the lane line,etc.The vision-based object detection is an important part of the perception of smart vehicles.Therefore,designing an effective object detection algorithm is particularly important for intelligent vehicles to obtain surrounding environmental information.In this paper,we will analyze the YOLO model,one of the real-time object detection algorithms from many aspects.Aiming at some shortcomings in the YOLO model,we combine the merits of other models to improve them.However,there are still some particularity in the application of the object detection algorithm in the actual traffic scenarios.Therefore,we propose an associated model based on conditional random fields to improve the accuracy of the algorithm.The main works of the thesis are as follows:1.We have thoroughly discussed the research progress of image recognition and object detection,and summarized the basic theories and related technologies of object detection,including convolutional neural networks,YOLO model,IOU and NMS.2.Aiming at the problems existing in the YOLO model,improvement schemes are proposed in combination with the desirable aspects in other models.The third and fourth layers of the original network are replaced by Inception structure,and a space pyramid pooling layer is added in front of the full connection layer of the original network.The advantage of doing that is to reduce the network parameters while widening and deepening the network,and it is further conducive to extracting features of different scales.Then,a comparative experiment was conducted on the PASCAL VOC 2007 data set to compare and analyze the effectiveness of the improved model.3.In view of the particularity of the object detection algorithm applied in actual traffic scenarios,that is,the vehicle target detection is based on the continuous video image,and the target will not suddenly appear or disappear,we propose an associated model based on conditional random fields.We improve the accuracy of the object detection by associating the historical detection data of the object in the video image sequence.Then,a comparative experiment was conducted on the KITTI dataset to compare and analyze the validity of the association model.