Research Of Object Detection Algorithm Based On Key Points

In recent years,with the development of deep learning and neural networks in computer vision.Target tracking,Human gesture recognition,Abnormal behavior detection,and Autopilot have developed rapidly.In the civilian field,high-precision and fast object detection algorithms can deployed on edge devices and achieve automatic parking,face recognition,abnormal behavior detection and etc through the video captured by camera;In the military field,object detection algorithms can applied to optical remote sensing images are helpful to determing the type and scale of hostile targets from input image.Attacking on targets precisely and maintain national security.The application of deep convolutional neural networks can improve image feature extraction ability than traditional object detection and improves the accuracy of object detection.The YOLO algorithms have both high accuracy and real-time performance.However,images downsampling by YOLOv3 network will compress the image to a smaller size,this will lose the context information of small objects.It is necessary to calculate anchor box manually before training start if you choose a new data sets.Calculating anchor box is time-consuming and CornerNet algorithm leaves out the step of calculating the anchor box.However,seeing a pair of corner points as the target feature will bringing a difficult problem that how to judge a pair of corner points detected are a pair from the same target actually.Refering the above problems,this paper proposes a object detection algorithm suitable for complex scenes based on the key points of the target and deep convolutional neural network.The main work of this paper is as follows:(1)Introduce the basic concepts of convolutional neural network and object detection algorithms.Then we introduce target detection algorithm based on key points.At last we decide use DLA-34 deep convolutional neural network for extracting objects key point features and achieve object detection mission based on CenterNet.(2)We establish our database for training based on the data from open source dataset though data augmentation and data enhancement.(3)Original CenterNet algorithm did not consider the shape and size of the target when extracting key points.The algorithm for extracting key points is simple so we propose a method to extract key point features considering the shape and size of the target.In the training process,we use the category of key points,the size of the target and the offset of the target center point as the loss to optimize the network parameters.Considering the background of the target in the training data set is complex,each image have multiple target and the positive and negative samples are imbalanced.In training process,we use focal loss instead of the traditional cross-entropy loss to reduce the weight of simple negative samples.We propse an end-to-end object detection algorithm.The input can be of any size so that we can incease accuracy without the image preprocessing.The key point extraction in this paper focuses on the shape and size of the target which is more robust.At the same time,the algorithm can achieve conversion from target to point,point to classification and location directly.The training process does not need anchor boxes non-maximum suppression so that we can shorten the time of network training and detection.