Research On Action Recognition And Salience Detection Based On Multiple Instance Learning

Both human action recognition and object salience detection are the hot researchtopicsinthefeldofcomputervisionforrecentyears. Humanactionrecognition,whichis the key technology in the application of video retrieval, video surveillance and in-telligent human-computer interaction, has extensive application prospect. But it is stillfull of challenge due to the problems of inter-person diferences, variations in humanaction and so on. Object salience detection, which aims at fnding the object regionswith vision salience in images, is of great help for object detection, image compres-sion and content-based image retrieval, and it has extensive application prospect andimportant research signifcance.In this thesis, we investigate the multiple instance learning algorithm and its cor-responding optimization algorithms, and we apply it to human action recognition invideos and object salience detection in images. We propose a human action recogni-tion approach based on key segment mining and an object salience detection approachbased on compounded supervised learning.Firstly, we propose a human action recognition approach based on key segmentmining. Based on multiple instance learning, our approach is able to fnd the locationand length of the most discriminative segment automatically for each action sample. Inaddition, a novel skeleton joint feature is proposed based on the3D coordinates of hu-man skeleton joints output from Kinect. It can efectively capture informative motionand shape cues of skeletons, and leads to a compact and discriminative representa-tion. The experimental results validate the efectiveness of the proposed human actionrecognition method and feature representation. Our method achieves the state-of-the-art performance on UCF-Kinect dataset, and shows superior accuracy than previous methods using only skeleton data on MSR Action3D dataset.We also propose an object salience detection approach based on compounded su-pervised learning. Through the combination of supervised learning and weakly su-pervised learning, our approach only needs to give precise annotation for part of thetraining data. For the rest images, we only need to give the bounding box of the ob-ject. Compared with supervised learning, our method reduces the annotation cost fortraining images; compared with weakly supervised learning, our method has higherprecision of salience detection. The experimental results demonstrate our approachcan efectively measure the tradeof between salience detection precision and cost forobject annotation. The results also show our approach can increase the precision ofobject salience detection via adding training images with only bounding box, whichdemonstrates a promising prospect for generalize salience detection to a large scaleapplication.