guest ::
| Home > Publications database > Two Methods for Detecting Pushing Behavior from Videos: A Psychological Rating System and a Deep Learning-based Approach |
| Home > Publications database > Two Methods for Detecting Pushing Behavior from Videos: A Psychological Rating System and a Deep Learning-based Approach |
| Conference Presentation (Invited) | FZJ-2021-05138 |
; ; ; ; ; ;
2021
Please use a persistent id in citations: http://hdl.handle.net/2128/29436
Abstract: In crowded entrances, some people try to be faster and therefore start pushing others. This pushing behaviorpotentially increases density, and decreases comfort as well safety of events. From research and practical perspectives, it is interesting to know where, why, and when pushing appears and, thereby, to understand the heterogeneity of movements in crowds. This paper presents two methods for identifying pushing in videos of crowds. The first one is a newly developed psychological rating system. It categorizes forward motion of people into four classes: 1) falling behind, 2) just walking, 3) mild pushing, and 4) strong pushing. The rating is performed by trained human observersusing the software PeTrack. This procedure allows to annotate individual behavior in every second, resulting in a high time resolution. However, this approach is time-consuming. The second method is an automated tool that can assist ad-hoc recognition of pushing behavior. We propose a novel deep learning-based technique that automatically detects pushing behavior scenarios from videos. In particular, we combine deep optical flow information with wheel visualization techniques to extract useful motion features from video sequences and generate a motion feature map between each two consecutive frames that visualizes: the motion speed, motion direction, spaces in crowd and interactions between pedestrians. Then, convolutional neural networks are used to extract the most relevant features (deep features) from these maps. Afterwards, additional supervised convolutional neural networks are used to automatically learn from the deep features to classify frames into pushing or non-pushing behavior classes. To evaluate this approach, we have conducted experiments using manually annotated videos by the first method. Results demonstrated a high congruence of both approaches and a promising performance in identifying pushing behavior from videos.
|
The record appears in these collections: |
PDF
Fulltext by OpenAccess repository