Research On Detection Of Vehicles By Adaboost Algorithm Based On GPU Acceleration

With the development of the study of computer science,the analysis of the intelligent video monitoring becomes a hot point.The core techniques of it are classifying and detecting of the target.The Adaboost(Adaptive Boosting)is one of the most widely used classify algorithms.It uses the same set of samples to train different weak classifiers,which finally be integrated to become a strong classifier which has much stronger classify ability.Some of the complicated Artificial Intelligence algorithms cannot be used on low-end equipment in engineering field.But Adaboost is not that complicated so it still can be used in lots of engineering fields.With the improvement of the technology,the resolution ratio of pictures and videos are improving,so the quantity of data is really big,which leads to the wasting of detecting time and the low efficiency.GPU(Graphic Processing Unit)was designed to be a graph deferring chip.But since its special hardware structure,the GPU has an outstanding ability of calculating.The specialists noticed that ability and tried to make use of it,especially in the big data area,which processes big amount of data.In order to optimize the traditional Adaboost algorithm and reduce the time of detecting,this essay makes use of GPU hardware to present the study and parallel improvement in detecting time.The main studies of this essay are as follows:(1)We analyzed the traditional Adaboost algorithm to see which parts are the most time-consuming,and discussed the possibility to be paralleled.We also made use of the CUDA(Compute Unified Device Architecture)and OpenCV platforms to make good use of the storage structure.Since the data transform is quite time-consuming,we reduced the transform as possible.We also measured the speed-up in different parts of the paralleled Adaboost algorithm.The result shows that all the parts can be accelerated.(2)Images with different resolutions are used to verify the influence of the image resolutions to the paralleled Adaboost algorithm.The result shows that higher the image resolution,better accelerating result we can get.(3)The speed-up version of the Adaboost detecting was used in the real monitoring environment.We used the paralleled Adaboost algorithm to detect the vehicles on the street and in the gas station.The result shows that the paralleled algorithm can get the speed-up ratio of 4 to 9.The resolution of monitor video has an influence on the speed-up ratio.The experiments showed that making use of the GPU and accelerating the Adaboost detecting of the target are feasible.It can also be used in the real monitoring environment.