Target Detection And Location Based On The Improved Faster RCNN Algorithm

As an important research branch in the field of computer vision,target detection technology requires to be able to accurately locate and classify objects,which provides a foundation for subsequent computer vision tasks.At present,there are still some defects in the operability of the practical application of this technology at home and abroad.In the operation process,it is difficult to accurately identify and the system application fluctuates greatly.Based on such problems,this paper analyzes the traditional target detection algorithm,deep learning target detection algorithm and binocular vision,proposes an improved algorithm,and carries out analysis through experiments to verify the accuracy and effectiveness of the improved algorithm.Firstly,this paper research on target detection algorithm based on Faster RCNN was carried out,and the network structure of target detection model was improved based on windows system and gpu server platform.By comparing YOLO algorithm,SSD algorithm and Faster RCNN algorithm,the target detection method based on Faster RCNN?Res net101 was determined,and the updated NMS algorithm was built to upgrade the candidate box,and the missed detection problem during the experiment was improved.Second,the COCO public data set and the target model in natural environment were respectively used to train and test the Faster RCNN network model before and after the improvement on the TensorFlow deep learning framework.The experimental results were compared and analyzed according to ap and fps and other indicators.The results show that the network model has a better recognition effect after optimization.The tensor board visual interface observation and evaluation were used to observe the convergence degree of total loss curve and classified loss curve,showing us the effect of the improved algorithm.Next,in order to solve the problem of large dimension and low accuracy of SIFT algorithm in traditional target detection,this paper proposes an improved method to reduce the traditional 128 dimension to 64 dimension.Moreover,RANSAC algorithm is selected to eliminate a large number of mismatched feature point pairs.It can be judged from the experimental results that the improved SIFT algorithm can increase the accuracy of feature extraction experiment in traditional target detection and reduce the running time.Through experimental comparison,the result of deep learning target detection algorithm is better than the traditional target detection SIFT algorithm.Finally,the target detection and location based on the improved Faster RCNN algorithm is studied.Experiments of binocular camera calibration and pole-line correction were carried out,and combined the improved Faster RCNN algorithm with the BM algorithm.The experimental results achieved better recognition,ranging effects and real-time performance.