| With the increasing number of cars per capita in China and the urbanization of the population,traffic safety and traffic congestion are becoming more and more prominent,and people’s demand for a safe and comfortable driving environment is becoming more and more urgent.Highperformance assisted driving system can well meet people’s needs,and real-time road environment perception is the prerequisite for the application of assisted driving system.In recent years,the performance of deep learning-based object detection algorithms in image processing has been greatly improved and widely used in various fields,but for scenarios such as assisted driving,which requires high real-time performance and has a complex and changing environment with many small targets,there are still defects such as low detection accuracy and poor real-time performance.Therefore,this paper researches on improving the accuracy and speed of the object detection algorithm,which is important to enhance the environment perception capability of the assisted driving system and promote its practical application.In this paper,based on the most widely used YOLOv3 algorithm,the basic convolutional structure,the feature extraction network and the structure of the neck part are improved as follows and experimentally verified on the autopilot dataset BDD100 K.（1）An efficient lightweight convolutional structure NFM module is proposed,which can greatly reduce the number of parameters and computation of convolutional operations compared with standard convolution.According to the characteristics of the NFM structure,we propose a method of gradually increasing the number of channels in the feature layer with the increase of the perceptual field,and redesign the feature extraction network.In addition,the depth-separable convolution is used to improve the structure of the neck part.The above lightweight improvements improve the speed of the model by 214.02% and compress the number of parameters by 90.38% at the cost of 2.2% decrease in m AP.（2）The size of the input feature map of the neck part is expanded according to the characteristics of complex road environment and too many small targets,which greatly improves the accuracy of the model at the cost of reducing the model speed.Based on the expanded size of the neck input feature map,a new neck structure RCFPN is proposed,which further improves the model m AP by 0.9% while increasing the model speed by 14.32% and compressing the number of parameters by 14.81%.These improvements result in a lightweight high-precision target detection model,RCFYOLOv3,which outperforms the original YOLOv3 by 9.3% in m AP,88.27% in runtime,and91.43% in number of parameters,which makes the model valuable in other fields,especially in embedded and other hardware resource-constrained fields.The lightweight convolutional structure NFM,the neck structure RCFPN and the new feature extraction network design method proposed in this paper provide new directions for the lightweighting of the object detection algorithm and the improvement of the detection accuracy,which are of good practical value for future model improvements. |
