Multi-timescale Trajectory Prediction for Abnormal Human Activity Detection

This repository will contain the implementation of the approach described in the paper,

@InProceedings{Rodrigues_2020_WACV,
author = {Rodrigues, Royston and Bhargava, Neha and Velmurugan, Rajbabu and Chaudhuri, Subhasis},
title = {Multi-timescale Trajectory Prediction for Abnormal Human Activity Detection},
booktitle = {The IEEE Winter Conference on Applications of Computer Vision (WACV)},
month = {March},
year = {2020}
} 

Our newly proposed dataset is now available for download. Kindly, contact me at royston.rodrigues@protonmail.com to request for a download link. Please cite our paper if you find our work or dataset useful for your research.

Project Updates

Oral Presentation at WACV 2020

IMAGE ALT TEXT HERE

Trajectory Predicition for Normal Samples

The proposed model can predict future human pose trajectories successfully for normal pedestrian activities. (Please reload the webpage if the gif files don’t appear in synchronization)

GitHub Logo GitHub Logo
Video Frames Input Trajectory, Our Predicition
GitHub Logo GitHub Logo
Input Trajectory, Ground Truth Input Trajectory, Ground Truth, Our Predicition

Trajectory Predicition for Abnormal Samples

The proposed model generates large deviations for the prediction of human pose trajectories corresponding to abnormal activities. (Please reload the webpage if the gif files don’t appear in synchronization)

GitHub Logo GitHub Logo
Video Frames Input Trajectory, Our Predicition
GitHub Logo GitHub Logo
Input Trajectory, Ground Truth Input Trajectory, Ground Truth, Our Predicition

IITB-Corridor Dataset

In order to benchmark existing techniques for the task of abnormal activity detection we introduce a new data-set, which consists of group activities such as protest, chasing, fighting, sudden running as well as single person activities such as hiding face, loitering, unattended baggage, carrying a suspicious object and cycling (in a pedestrian area). We believe such a dataset will motivate investigations for human activity analysis to consider single human or multi-human interaction. We name it the IITB-Corridor dataset. IITB-Corridor is a large scale surveillance dataset with 4,83,566 frames and will be made available for free for research purpose.

GitHub Logo GitHub Logo GitHub Logo
Protest Sudden Running Fighting
GitHub Logo GitHub Logo GitHub Logo
Unattended Baggage Chasing Suspicious Object
GitHub Logo GitHub Logo GitHub Logo
Hiding Face Playing with Ball Loitering
GitHub Logo GitHub Logo GitHub Logo
Cycling Standing Walking

Download Information for IITB-Corridor Dataset

IITB-Corridor Dataset Download Link :-

1.) Train Set :- https://drive.google.com/file/d/1HZZjINXIgWnq1FYuVTTBsfJiWsXy1uU5/view?usp=sharing

2.) Test Set :- https://drive.google.com/open?id=1F0m6kRcVKAvDIhGLgOY4QJ-oFdTmkzPI

Password for both the files : tb7bdEZE

Please note, this dataset is made available for research purposes only.

Results

To compare with existing approaches, we use frame level AUC as the evaluating criteria. Our proposed model outperforms existing methods. The comparison is given in the below table.

Method HR-Avenue HR-ShanghaiTech ShanghaiTech IITB-Corridor
Liu et al. (CVPR-18) 86.20 72.70 72.80 64.65
Morais et al. (CVPR-19) 86.30 75.40 73.40 64.27
Ours 88.33 77.04 76.03 67.12

Contact

Feel free to contact me at royston.rodrigues@protonmail.com for queries regarding this work.