Fight detection from video is a task with direct application in surveillance scenarios like prison cells, yards, mental institutions, etc. Work in this task is growing, although only a few datasets for ght detection currently exist. The Violent Scene Detection task of the MediaEval initiative offers a practical challenge for detecting violent video clips in movies. In this working notes paper we will brie y describe our method for ght detection. This method has been used to detect ghts within the above-mentioned violent scene detection task. Inspired by results that suggest that kinematic features alone are discriminative for at least some actions, our method uses extreme acceleration patterns as the main feature. These extreme accelerations are e ciently estimated by applying the Radon transform to the power spectrum of consecutive frames.
In the last years, the problem of human action recognition
from video has become tractable by using computer vision
techniques [
The presence of large accelerations is key in the task of
ght detection ([
Motion blur entails a shift in image content towards low frequencies. Such behaviour allows building an e cient acceleration estimator for video. Our proposed method works with sequences of 50 frames; therefore we need to divide the shots in 50-frame clips. First, we compute the power spectrum of two consecutive frames. It can be shown that, when there is a sudden motion between two consecutive frames, the power spectrum image of will depict an ellipse. The orientation of the ellipse is perpendicular to the motion direction, the frequencies outside the ellipse being attenuated. Most importantly, the eccentricity of this ellipse is dependent on the acceleration. Basically, the proposed method aims at detecting the sudden presence of such ellipse.
Our objective is then to detect such ellipse and estimate its eccentricity, which represents the magnitude of the acceleration. Ellipse detection can be reliably performed using the Radon transform, which provides image projections along lines with di erent orientations.
For each pair of consecutive frames, we compute the power spectrum using the 2D Fast Fourier Transform (in order to avoid edge e ects, a Hanning window is applied before computing the FFT). Let us call these spectra images Pi 1 and Pi. These images are divided, i.e. C = Pi=Pi 1.
When there is no change between the two frames, the power spectra will be equal and C will have a constant value. When motion has occurred, an ellipse will appear in C. Ellipse detection can be reliably performed using the Radon transform.
After applying the Radon transform to image C, its vertical maximum projection vector is obtained and normalized (to maximum value 1). Next when there is an ellipse, this vector will show a sharp peak, representing the major axis of the ellipse. The kurtosis of this vector is therefore taken as an estimation of the acceleration.
Note that kurtosis alone cannot be used as a measure, since it is obtained from a normalized vector (i.e. it is dimensionless). Thus, the average value per pixel P of image C is also computed, taken as an additional feature. Without it, any two frames could lead to high kurtosis even without signi cant motion.
Deceleration was also considered as an additional feature, and it can be obtained by reversing the order to consecutive frames and applying the same algorithm explained above. For every short clip, we compute histograms of these features, so that acceleration/deceleration patterns can be used for discrimination.
Once we have features for every clip, for classi cation we use two di erent classi ers: the well-known K-Nearest Neighbours and Support Vector Machine (SVM) with a lineal kernel.
The method described requires training clips that contain ghts. Thus, when ght sequences are given, they may have to be rst evaluated for ght subsequences.
Sometimes, within a violent segment non- ght clips may appear. For each violent segment (as provided by MediaEval organizers), we manually removed clips without ghting action. Moreover, random clips of 50 consecutive frames, taken outside the violent segments, were selected for the non- ght class. We trained the two classi ers for the ght concept. Then we apply the two trained classi ers to every 50 consecutive frames of the test set.
We submitted 2 runs and the details of performances are
as follows. Table 1 reports AED [
Based on the observation that kinematic information may su ce for human perception of various actions, in this work a novel ght detection method is proposed which uses extreme acceleration patterns as the main discriminating feature.
In experiments with other datasets we obtained accuracies above 90%, and processing times of a few milliseconds per frame. The results on the MediaEval dataset are, however, very poor. We suppose it could be due to the test ground truth (used by the organizers to obtain the performance measures). The category 'violence' is more general because includes violent scenes that could include explosions, shots, car chases, ghts, etc. Although we had ' ght' labelled training videos, that label was not available for test videos. What is more, the de nition of 'violence' in MediaEval is "physical violence or accident resulting in human injury or pain", so we were able to detect only part of the violence: ghts.
On the other hand, there are a number of practical aspects that are not taken into account. In many surveillance scenarios, for example, we do not have access to color images and audio, and the typical forms of violence are ghts and vandalism, instead of explosions, car chases, etc. The processing power needed for running detection algorithms is also an important issue in those applications. 5.
This work has been supported by Project TIN2011-24367 from Spain's Ministerio de Econom a y Competitividad.
The authors also thank the MediaEval organizers for inviting us to participate. 6.