The Research On Multi-target Detection And Recognition Method Based On Improved Darknet Framework

With the aggravation of urban traffic congestion,in order to solve the low detection rate and poor robustness of the traditional vehicles detection method,which needs to extract different features as the scene changes,a vehicle multi-target detection and recognition method based on Darknet framework was proposed.On the basis of YOLO V2 algorithm,the YOLO-voc network was improved according to the change of the target scene and the traffic flows.The classification training model was obtained base on the ImageNet data and fine-tuning technology.Then parameters were adjusted according to the training results and the vehicle characteristics.Finally,the YOLO-vocRV network model was obtained which is more suitable for road vehicles detection.In order to verify the validity and completeness,the experiment was carried out for different vehicle flow states.The research content mainly includes:(1)A multi-vehicle detection method was adopted,which bases on YOLO 9000 under Darknet framework.The YOLO9000 structure was improved according to the the training results and vehicle target characteristics,the algorithm parameters were adjusted.Finally,the YOLO9000-md network model was obtained,which is more suitable for current road video vehicles detection.And the vehicles under video was tested.The experimental results show that,the recall rate of vehicle multi-vehicle detection method based on improved YOLO9000-md model reaches 96.15%,which has certain validity.(2)A multi-target detection method was proposed based on the YOLO v2 algorithm model under the Darknet deep learning framework.The experiment was carried out on different vehicle flow states of different traffic densities.And the improved method was compared with the YOLO-voc,YOLO9000 and YOLO v3 models.The results show that the accuracy and recall rate of YOLO-vocRV model are all gathered at 0.95,so the recall rate of loss is obviously smaller under the condition of obtaining better accuracy,which achieves a good compromise.After training with mixed samples,the detection rate of vehicles multi-target detection method based on YOLO-vocRV model can reach 99.11% in free flow state,97.852% in synchronous flow state,97.311% in blocking flow state.It has a small false detection rate and good robustness.(3)The research on vehicles type recognition based on improved YOLO-vocRVmodel.First,using the Softmax classifier,the effectiveness of the improved model was verified,and the classification model of 60,000 iterations was obtained.Then,the improved model was compared with the typical model YOLO9000 and YOLO-voc.The experimental results show that the improved model can obtain better mAP for different vehicles,and the average mAP of different vehicles type can reach 88.24%.Finally,the improved model was trained with increased data samples,the average accuracy in detecting single target and multi-target are 92.21% and 89.44%.The detection effect is good and universal.There are 44 figures,8 tables and 58 references in this paper.