High-performance 3D Object Recognition Research

Object recognition algorithm has high complexity, Compute-intensive and slow response feathers, A high recognition rate algorithm demands heavily calculation, then it leads to a slow response. This situation can be alleviated by increasing the computing power and improving algorithms. In this paper, by proposing multi-level of details recognition algorithm, using heterogeneous computing architecture consisting of GPU and multi-way multi-core CPU, and by using CUDA interface to achieve massively parallel image processing, OpenCV for auxiliary calculations, real-time object recognition was efficiently implemented.First, multi-level of details recognition algorithm consisted of BPNN Strong Classifier and object's feature-database is designed. BP neural network can efficiently implement in parallel, it is strong classifier, beneficial to establish object's feature-database, yet the popular Haar Cascade Classifier is not appropriate to the establishment of object's feature-database and difficult to parallel implementationZbesidesZit can be found that the weak classifier complexity can great impact on the Ada Boost algorithm performance, and Haar Cascade Classifier's weak classifier is lower complexity.Secondly, image processing algorithms such as histogram equalization and etc were implemented in parallel at the GPU, and the forward process of BP neural networks was calculated through GPU, then logical decision for multi-level of details recognition was implemented by using CPU, object's feature-database was made in accordance with data-driven system designed style. Finally, designed and implemented image processing series configuration and scheduling module. That proposed an efficient, based on heterogeneous computing systems, flexible framework for object recognition.Through software implementation, results showed that by using of heterogeneous computing architecture and multi-level of details recognition algorithm, target can be efficiently identified.