Research On The Key Technology Of Object Detection Based On Binocular Stereo Vision

Object detection has played an increasingly important role with the development of artificial intelligent applications,such as self-driving cars,robot navigation.Aiming at the problems of low detection accuracy and slow detection speed,the key technology of object detection based on binocular stereo vision is proposed in this paper.The main contents and innovative achievements are as follows:(1)By analyzing and summarizing the basic framework of the object detection system and the related algorithms of each module,the CNN classifier based on region proposal method is proposed in this paper.Traditional object detection algorithms are mostly based on the combination of sliding window paradigm and manual feature extraction.The sliding window algorithm needs to scan every position and scale of the image,which limits the detection rate of the system,and manual feature extraction has a direct impact on the performance of the object detection system.(2)Aiming at the problem of poor performance of traditional region proposal method,the method is improved in this paper by using Bayesian integration with the goal of using as few region proposals as possible to cover as many objects as possible.Implementation method can be roughly divided into three steps: firstly,a large number of candidate proposals are generated by traditional region proposal methods;secondly,several geometrical features and depth informed features are extracted from each proposal;finally,using Bayesian to integrate features and computing the probability of every proposal which is predicted to be a positive sample.The probability is used as score of each proposal and the required number of proposals can be selected according to score.(3)For the impact on the performance of algorithm caused by depth calculation and ground estimation,we make a lot of statistics of histogram distribution for the features extracted from bounding boxes with manual annotation and discover that the real object size and the height above ground obey normal distribution.This paper proposes to integrate these two features by Bayesian framework after MLE estimation.(4)This dissertation studies the implementation of object detection system based on deep learning.Using Fast R-CNN as the basic framework,we apply improved region proposal methods to generate a set of proposals and select 1000 proposals per image with highest scores as the input of RoI Pooling layer.Then,all proposals are classified and regressed.Through the comparative analysis of experiments,the detection accuracy of the enhanced system has been greatly improved,with an average increase of over 15%,which verifies the effectiveness of the scheme.