Application Of Deep Learning In Object Detection

Object detection technology has always been a hot topic in the field of computer vision.With the continuous development of deep learning,object detection technology based on deep learning has also achieved good results.Nowadays object detection technology has been widely used in various fields.With the improvement of people's living standards,the number of cars in transportation has always been increasing.While vehicles bring convenience to people,they also bring a series of hidden traffic safety hazards,which seriously threaten people ' s lives,therefore,it is esential to monitor vehicles on the road.Here this article takes the vehicles on the road as the detection object,and uses the object detection method based on deep learning for research.Firstly,this article compares the one-stage and two-stage object detection algorithms in deep learning.Under the trade-off between detection accuracy and real-time performance,the one-stage object detection algorithm YOLOv3 is selected as the basic algorithm of this research object,and the vehicle data set is used to carry out the YOLOv3 algorithm.Aiming at the problem of the YOLOv3 alogrithm with complex calculations and large par ameters during the training process the lightweight network Mobile Net V3 is used to improve the original Dark Net53 network,reducing the computational complexity and the number of parameters of the entire model,and improving the detection of the model speed.Then,in order to further improve the detection accuracy,the Soft NMS alogrithm is used to improve YOLOv3 for the problem that the YOLOv3 alogrithm cannot handle overlapping targets very well and affects the detection effect.Aiming at the problem that the anchor frame size used in yolov3 is not suitable for this vehicle detection,kmean ++algorithm is used to re-cluster the anchor frame in the vehicle data set,so as to improve the detection accuracy of the model;Due to the limitation of experimental hardware conditions,the error of Batch normalization in yolov3 is large when the batch is small.Group normalization is introduced to improve the algorithm.Through the use of lightweight network and the above improvement scheme to improve the detection accuracy,experimental verification shows that the final detection accuracy of the model has been increased from 81.3% to 84.4%,the frame rate has been increased from the original 41 FPS to 62 FPS under GPU,and the model weight file size has been reduced from the oreginal 236 MB to 32 MB.Finally,a vehicle detection system was designed and developed,and the improved YOLOv3 algorithm was applied to the vehicle detection system to complete the automatic vehicle detection function for the pictures and video streams input to the system.