| Image classification and object tracking have always been the core issue of computer vision.In recent years the convolution neural network(CNN)has achieved better and better results in image classification.Also,in object tracking area,the Tracking-LearningDetection(TLD)algorithm made the tracking process no longer depends on the pretrained model.In the same time,mobile devices like phones and tablets are getting so powerful that people start to think how to apply these amazing computer vision algorithms on mobile devices.For the CNN,we first implemented the forward process on a mobile phone with the deep learning framework MXNet.We analyzed the execution time of different layers in the network and found the convolution layers take up more than 70% of all.We offload the computing work of these layers to the other device in the phone – GPU,which is more suitable for compute-intensive tasks.Based on the structure of mobile GPU and the work load of the task,we have taken several methods to improve the performance,like dividing the task,using on-chip memory,using wider data types and so on.In the final test of our high-performance implementation,we accelerate the matrix multiplication in convolution layer by 16 times,and the time of forward process also gets a speed-up of about 2.1.With the Consideration of the bottleneck of TLD lies in the detection module,which is consist of three classifiers,the patch variance classifier,the assemble classifier and the nearest neighbor classifier,we offload this part of task to the mobile GPU.We also made some improvements to gain better performance.In the tests,our high-performance implementation of TLD on mobile GPU successfully reduced the time of tracking an object in a single frame to only a half.This thesis tries to study how to make better performance on mobile GPUs in general parallel computing through implementing the CNN and TLD algorithm on mobile devices,and the experiment results have shown we have made the ideal acceleration effect. |
