Research On Detection Algorithm Of Moving Multi-target

With the deepening of research,target detection and recognition technology based on deep learning has already entered the engineering field.However,currently,the detection framework based on single frame is widely used,which can be applied to both single frame image task and multi-frame video task.Video detection and recognition task has a characteristic,is the video is made up of many frame,front and rear frames in time,there was a link is the same goal in the change of adjacent frame has the movement information,although the existence of this motion information will cause the focus blur,motion blur,but if can fully use this motion information,also can improve the target detection and recognition ability.Therefore,the frame of a single frame cannot satisfy the task of multiple frames.How to make use of the time and motion information of adjacent frames is the research direction of this paper.In particular,the data set used in this paper is the video data shot on the UAV,which makes the detection and recognition task have to face the target complexity and other problems.At the same time,the motion of the UAV leads to the picture shaking and blurry.The technical problem to be solved in this paper is to study how to use the time motion information between adjacent frames to design an algorithm with better performance than the single frame,so as to avoid the disadvantages of the single frame.This paper adopts two different ideas to design the algorithm.The main work includes the following aspects:Firstly,the MG-SSD(Memory Guided Single Shot Multi Box Detector)target detection and recognition algorithm based on the memory module and interleaved model is constructed.Since the feature representations on adjacent frames are often similar,it is the core concept of the MG-SSD algorithm to use past features to improve the current detection and recognition results in a memorized manner.It combines the LSTM from the recurrent neural network with the convolutional layer in target detection to form the memory module in this detection and recognition framework.On this basis,the Bottleneck design is used to improve the performance of the LTSM layer,the Shuffle Net unit is used to replace the Res Net unit in the general deep residual network to reduce the consumption of computing resources,and the deconvolution module is used to improve the network's ability to detect small and medium targets.The design of the interleaved model fully embodies the benefits of using time information,optimizes the memory retained by the network with accuracy,and improves the speed of the network's capabilities.Secondly,FG-FCN(Flow Guided Fully Convolutional Networks)target detection and recognition algorithm based on optical flow feature propagation and aggregation is designed.Unlike the MG-SSD algorithm,which uses a memory module to retain some features,FG-FCN uses an optical flow network to predict the optical flow field,thereby improving the quality of detection and recognition by means of feature propagation and aggregation.Among them,sparse recursive feature aggregation is used between key frames and key frames,and spatial local feature updates are used between key frames and non-key frames.Since the performance of the optical flow network also affects the entire algorithm,hollow convolution and channel attention mechanisms are used to improve the optical flow network.In addition,there are design key frame scheduling algorithms,and R-FCN with deformable convolution as the detection network.Finally,the difference between the accuracy and peed of the two algorithms designed in this paper under different hyperparameters is compared on the data set taken by the drone,and compared with other algorithms.It is found that the accuracy of FG-FCN algorithm based on optical flow feature propagation and aggregation is higher than that of MG-SSD algorithm based on memory module and interleaved model in the case of small hyperparameters,but after the hyperparameters gradually increase,MG-SSD algorithm performance is gradually better than the FG-FCN algorithm,and in general,the inference speed of the MG-SSD algorithm is faster than the FG-FCN algorithm.At the same time,in order to reflect the generalization ability of the algorithm designed in this paper,through testing on self-collected data,it shows that the algorithm of this paper is universal and meets the needs.