Research On Visual Target Tracking Algorithm In Complex Scenes

With the continuous development of target recognition and tracking technology,deep learning has shined in the field of target recognition and tracking,and is widely used in various fields.With the continuous development of deep learning technology at China and abroad,various algorithms have emerged one after another and have been widely used in the field of target recognition and tracking.The target detection and tracking algorithm based on deep learning,this can quickly process information and extract the elements related to target detection for learning and classification,and improve the detection efficiency.With the continuous improvement of image resolution,how to effectively extract the detailed information in the image is particularly important in algorithm research,and it also brings new challenges to the development of target recognition and tracking technology.The thesis improves the existing target detection and tracking algorithms,and its main innovations are as follows:1?Aiming at the problem of the loss of details caused by long-distance small targets when the picture is enlarged and stretched,this thesis proposes an improved Lap SRN reconstruction super-resolution algorithm,that is,the bicubic interpolation method is introduced into the Lap SRN algorithm.The algorithm makes full use of the characteristics of bicubic interpolation that can effectively smooth the edges of the image.The simulation shows that the improved Lap SRN reconstruction super-resolution algorithm can accurately restore the details of the image when the image is magnified by 2 times,thereby alleviating the distortion caused by the image magnification.2?Aiming at small targets and occlusions in vehicle detection,this thesis improves the traditional YOLOv3 algorithm.The output of the second residual block is up-sampled with the down-sampled output of the third layer for feature fusion.The output of the five layers is fused with the fourth layer through two up-sampling,and the SPP module is added for multi-scale fusion.The simulation shows that the improved algorithm designed in this thesis has better ability to recognize small targets and occlude targets than the YOLOv3 algorithm,and the improved algorithm can effectively reduce the missed detection.3 ? Combining the improved YOLOv3 algorithm with the target tracking Deep-sort algorithm,that is,using the improved YOLOv3 algorithm to replace the target detection algorithm in the Deep-sort algorithm.Using the Kalman filter algorithm to advance the prediction output,and then the Hungarian algorithm matches the predicted target tracking frame with the detection frame.Finally this algorithm improves the accuracy of vehicle detection for long-distance small targets and occlusion situations.The simulation shows that the number of successfully tracked vehicles has increased after the improvement,and the loss of tracked targets has improved.