Research On Object Detection Based On Machine Vision Algorithm

Object detection based on machine vision is the military, civil and industrial production often faced with a very important issue. It is mainly through the computer to process the data, and then obtained from different sensors data object detection, tracking and testing, which involves multi-disciplinary, requires the use of different techniques and expertise. However, until now, there is no one way to solve the problem of detection of all objects, nor any one object tracking and detection systems can solve all the problems. Thesis research on the analysis of object detection algorithm, Hough transform algorithm and neural network research status, based on the different types of objects carried detection algorithm.For an existing line detection positioning accuracy is low, prone to multiple single-edge response and other issues, the paper presents a Canny operator based on improved Hough transform object detection algorithm based on segments from the common characteristics of the object edge detection operator child to find the best overall performance of operators, and is made accumulator unit subdivision derivation. The algorithm can effectively remove the noise interference, to solve the calculation accuracy and calculation of optimal matching between the speed problem, and effectively solve the multi-peak detection issues and false peak problem and improve the robustness of the line detection.For rigid object by itself and outside interference will not produce self-occlusion, deformation, shadow features, as well as existing features moments in the rigid object detection problem of low accuracy, from Hu moment invariants constructed ten invariants Combination moment, use notation and shape representation apparent Combination with a combination of these ten invariant moments to describe the features of an object. Paper proposes a rigid object based on neural network detection algorithm, this paper presents a given based on BP neural network classifier, in determining the input and output of the network after the practical application select the appropriate model structure parameters training, the algorithm error analysis and efficiency analysis, and analysis of some parameters on the network generalization ability of the classifier has better real-time and detection rates. Object pose for pedestrians complexity, diversity, pedestrian objects texture diversity and vulnerable to outside interference conditions particularity, in contrast to the existing class Haar features, Shapelet features and SIFT features, the design of a pair angular features to extend the class Haar features and uses approximation method for rapid calculation of the integral image mixed eigenvalues of the characteristic features while maintaining the scalar computing speed advantage, but also improves the ability to describe. This paper presents an AdaBoost algorithm based on hybrid features pedestrian detection algorithm, based on strong and weak classifiers characteristics cascade classifier constructed tree, during positive and negative sample detection speed, low time complexity of the algorithm, Real-time detection effect.These three different types of object detection algorithms are based on the vehicle tachograph to access to environmental information, the characteristics of the original algorithm and describe the means proposed to improve the algorithm and build new features. After comparing test results found that a variety of algorithms, constructed by the above method improved algorithm can be appropriately applied, the detection process adaptability, thus contributing to the object detection based on machine vision algorithm.