Research On High Precision Object Detection Technology Based On Deep Learning

With the continuous development of computer vision technology,it has been widely used in many fields such as intelligent city,automatic driving,security monitoring and so on.Target detection,as the core module of these fields,will directly affect the quality of the final results.Efficient and robust target detection algorithm has always been a hot and difficult issue in the field of computer vision.At present,target detection based on deep learning has become a research trend and is becoming more and more popular.This kind of algorithm learns a lot of labeled data independently to acquire the corresponding features,generates a specific detection model by means of supervisory training,and finally uses the above model to detect the target of unlabeled images,and obtains a significantly better detection result than the traditional algorithm.However,with the continuous development of deep learning technology,the requirement of training sample data is increasing gradually.The traditional manual labeling is not only time-consuming and labor-consuming,but also inefficient in accuracy.To a certain extent,it restricts the application of deep learning method in the field of target detection.Therefore,how to achieve high precision and fast automatic data annotation becomes the key to improve the efficiency of target detection algorithm.Based on this,this paper focuses on the problem that the precision of model-based automatic annotation method is not high,proposes a high-precision target annotation and detection technology based on deep learning,and tests it on the HIKVISION detection platform,and achieves good results.The specific work of this paper includes the following parts:1)Aiming at the current target detection model,including single stage target detection model and two stages target detection model,the advantages and disadvantages of various models in target detection accuracy are analyzed.2)Based on the characteristics of target detection and labeling tool,FPPI-Recall is used as the evaluation index of target detection and labeling tool,and Cascade R-CNN target detection algorithm is applied to HIKVISION labeling tool.The experimental results show that when IoU = 0.5 and FPPI = 0.3,the tagged FPPI-Recall index Recall parameter increases by 7.3 percentage points to 0.6118.3)The shortcomings and potential advantages of Cascade R-CNN target detection algorithm are analyzed.The original model is optimized from the aspects of sampling and alignment of deep learning model,multi-scale model training and model tailoring,and the optimized model is applied to the annotation tool.In summary,based on the key technology research of target detection and calibration algorithm,this paper studies and designs high-precision target detection technology,and completes the experimental verification.The development of calibration software based on deep learning and computer vision has been carried out.State of art calibration algorithm of accuracy and accuracy has been successfully realized.