Spatio-temporal human action detection and instance segmentation in videos

With an exponential growth in the number of video capturing devices and digital video content, automatic video understanding is now at the forefront of computer vision research. This thesis presents a series of models for automatic human action detection in videos and also addresses the space-time action instance segmentation problem. Both action detection and instance segmentation play vital roles in video understanding. Firstly, we propose a novel human action detection approach based on a frame-level deep feature representation combined with a two-pass dynamic programming approach. The method obtains a frame-level action representation by leveraging recent advances in deep learning based action recognition and object detection methods. To combine the the complementary appearance and motion cues, we introduce a new fusion technique which signicantly improves the detection performance. Further, we cast the temporal action detection as two energy optimisation problems which are solved using Viterbi algorithm. Exploiting a video-level representation further allows the network to learn the inter-frame temporal correspondence between action regions and it is bound to be a more optimal solution to the action detection problem than a frame-level representation. Secondly, we propose a novel deep network architecture which learns a video-level action representation by classifying and regressing 3D region proposals spanning two successive video frames. The proposed model is end-to-end trainable and can be jointly optimised for both proposal generation and action detection objectives in a single training step. We name our new network as \AMTnet" (Action Micro-Tube regression Network). We further extend the AMTnet model by incorporating optical ow features to encode motion patterns of actions. Finally, we address the problem of action instance segmentation in which multiple concurrent actions of the same class may be segmented out of an image sequence. By taking advantage of recent work on action foreground-background segmentation, we are able to associate each action tube with class-specic segmentations. We demonstrate the performance of our proposed models on challenging action detection benchmarks achieving new state-of-the-art results across the board and signicantly increasing detection speed at test time.

Grantor dc:publisher

Oxford Brookes University

Year dc:date

2018

Author dc:creator

• Saha, Suman

Contributors dc:contributor

• Cuzzolin, Fabio
• Crook, Nigel
• Olde Scheper, Tjeerd

Statement dc:rights

• All rights reserved

Language dc:language

en

DOI dc:identifier

https://doi.org/10.24384/kqtr-e820

OAI identifier oai:identifier

tle:bdee83eb-bb5c-4a68-afc9-45bde19510e9:d6bd9758-527a-46cd-bfe2-c433766e8fca:1