Research On Improvement Of Target Detection Algorithm Based On Depth Feature

Object detection is a key technology in the fields of driverless and face recognition.The traditional object detection algorithm generate the candidate object boxes first,then extract the handcraft features,and train the traditional classifier.In these methods,the ability of handcraft features expression is not strong,and the number of candidate object boxes is too large,which results in a low detection precision and slow computing speed.In recent years,deep learning based object detection research has been in an active state.The method based on region proposals and regression method in deep learning greatly improve the accuracy and speed of object detection.But there are also some problems such as poor detection effect on small object,limited expression ability of feature maps.Based on the existing problems,this paper proposes an improved object detection method based on Inception block and residual structure,and a multi-layer fusion networks.In the last we implement a real-time object detection system.The detailed work of this paper is as follows:1)We introduce the Inception block on the existing object detection frameworks,so as to get the receptive field with different scales around the objects,and enhance the effect of detection and recognition.In addition,in order to train a deeper networks and keep the object information as possible,we add a residual structure to the embedded module.Finally,in order to further improve the effectiveness of object detection,we analyze the shortcomings of the non maximum suppression algorithm,and proposes an improved algorithm.2)In order to improve the expressive of feature maps,we fusion multi-layers as a new object detector.We explore the best way of fusion,and experiment on the public dataset.In addition,in order to further improve the effectiveness of object detection,we analyze the existing problems in the prediction of the object boxes design,and proposes a priori object boxes design method based on cluster.3)We based on the proposed improvement method,implement an object detection system.This system realizes online and offline detection patterns.The online detection means real-time detection of the real scene through a camera device.The offline detection means that the user can use this system to detect local videos.