Research On Multi-object Tracking Algorithm Based On Convolutional Neural Network

Multi-target tracking is an important branch of computer vision.Both academically and commercially have important and tremendous value.In the field of defense,multitarget tracking is one of the key technologies for drones to complete reconnaissance and strike missions.In the field of monitoring,multi-target tracking can analyze the video,which can effectively reduce the burden of the police in our country and greatly improve work efficiency.In addition,multi-target tracking is one of the key technologies in unmanned driving and intelligent robots.The current multi-target tracking algorithm is mainly based on detection,the quality of the detection algorithm will directly affect the effectiveness of the tracking algorithm.When there are missed or false detections,it will affect the multi-target tracking algorithm.In addition,occlusion between targets,camera movement,target deformation,etc.are all problems that multi-target tracking algorithms need to solve.This paper proposes a multitarget tracking algorithm based on convolutional neural networks.The specific work and innovations are as follows:1.Sort out the common target detection algorithms,mainly introduce the target detection algorithms of YOLO series and R-CNN series,and compare the similarities and differences of these algorithms.2.Sort out the traditional multi-target tracking algorithm and the multi-target tracking algorithm based on deep learning.For the traditional multi-target tracking algorithm,the methods of multi-hypothesis tracking and structural constraints are studied.For deep learning-based algorithms,methods such as deep association measurement and identity re-recognition are studied.3.The multi-target tracking algorithm in this paper uses the Twin Area Recommendation Network(Siamese RPN)to generate new candidate targets.Use the target being tracked and the detection target of the current image sequence as template targets to generate new candidate targets.This method can improve the existing detection results,and has better robustness to the missed detection of the target.4.Use R-FCN to re-score the candidate targets and eliminate redundant candidate targets.Use the R-FCN network to re-score the candidate targets.After scoring all targets,a non-maximum suppression algorithm is used to eliminate redundant targets.By processing the target by this method,better candidate targets can be retained,and the original false detection targets can be cleared at the same time.Can effectively improve the detection results,and thereby improve the tracking effect.