| In the field of computer vision,object detection has always been a hot field in industrial applications.It involves the combination of object classification and object localization in scenes,such as face recognition,license plate recognition,pedestrian detection,and so on.With the continuous improvement of hardware computing power,deep learning based on the convolution neural network,which is applied by object detection,has been developed rapidly.And the real application of object detection needs a high level not only in the hardware aspect,but also in the algorithm speed,i.e.real-time is achieved.Therefore the accuracy and speed of the YOLOv3 algorithm in the deep learning are improved.The practicability of Jetson TX1 as the core embedded platform is studied.Finally,the object detection is realized accurately,efficiently and quickly.The specific work is as follows:First,in order to solve the problem that object detection requires high speed of the algorithm in practical application,YOLOv3 algorithm is improved by the ideas of ROI extraction,a new method of ROI extraction—NEW-ROI is proposed in the shallow layer network of YOLOv3 algorithm.Then the channel selection module is added on this basis.Finally the purpose of removing the region of non-interest(RONI)is achieved.Thereby the calculation amount is reduced and the acceleration effect is realized.Secondly,in order to solve the problem of error caused by the scale of the actual detected objects,the hourglass pyramid is embedded network into the Darknet-53 network of YOLOv3 algorithm.Not only the scale number of the output characteristics is increased,but also the number of network layers is deepened.Finally,the improved YOLOv3 algorithm was transplanted to the Jetson TX1 development board.The m AP of 70.6% is achieved at 21.1FPS.The experimental results show that the improved YOLOv3 algorithm could better detect multi-scale objects on the hardware and achieve the advantages of fast detection speed and high accuracy. |
PDF Full Text Request