Research On Object Detection And Multiple Object Tracking Methods Based On Lightweight Networks

Object Detection is one of the classic tasks in the field of computer vision.It is a basic visual task and plays an important role in image understanding.Image object detection belongs to the problem of region classification,which purpose is to locate and accurately classify the region of interest in the image.Multiple Object Tracking is another classic task in computer vision.Compared with the image object detection,the multiple object tracking is the continuation and extension of the image object detection,which input is video.Multiple Object Tracking not only classifies and locates the object in the frame,but also needs to distinguish the object by instance.After that,the same object between frames in the video is associated,and the trajectory of each object is output.In recent years,Deep Learning has developed rapidly in the field of computer vision.Compared with traditional image processing algorithms,convolutional neural network can efficiently extract image features.Based on this,convolutional neural network also provides a new research idea for object detection and multiple object tracking.With the need for industrial application,a series of lightweight network architectures have been developed.However,it is difficult for existing lightweight networks to achieve a good balance between speed and accuracy in object detection.Secondly,in the feature fusion stage,the existing lightweight object detection algorithms are deficient in utilizing deep semantic information.Based on these problems,this thesis conducts the following research:(1)In this thesis,we propose an efficient object detection network ECDET.Through the lightweight FPN for feature fusion,the model as a whole is lightweight.Experiments on Pascal VOC and MS COCO public datasets demonstrate that the proposed method achieves better results in the balance between detection accuracy and efficiency.(2)On the lightweight object detection algorithm for deep semantic information using the problem of insufficient,we propose a real-time object detection based on residual attention to guide network RSANet in this thesis.Through the guidance of attention mechanism,the semantic information from the deep layer in the form of residual can help shallow features gain stronger discriminant ability.Significant accuracy improvements were achieved on both Pascal VOC and MS COCO datasets without a large number of extra computations.(3)Based on the current mainstream of multiple object tracking algorithm based on detection of the problem of low speed on the CPU inference,in this thesis,we employ the ideas of the lightweight,migrate from the object detection multiple object tracking task,and propose an method Fast MOT,which makes the proposed method on the CPU can greatly accelerate the inference speed.