Recognizing Human Actions by Using Spatio-temporal Motion Descriptors - Advanced Concepts for Intelligent Vision Systems - page 367

Our method is closely related to the work of [12] and [14]. The main difference
is that instead of image gradients we calculate the gradients of the difference of
consecutive video frames, therefore our descriptor contains motion information
only. Moreover, we evaluate two special arrangements for the quantization of
the directions. An action is represented by these features in a fixed size spatio-
temporal rectangular cuboid, and this single 3-D descriptor (on single scales)
can be used for recognizing actions on the multi-scale representation of the input
video segments.

3Propo edM thod

The proposed method is based on the original HOG-based human detector of
[13], however we extended it with a third dimension. That is we calculate our
features in spatio-temporal rectangular cuboids. Two different motion features
are extracted: histograms of optical flow directions ( HFD ) and histograms of
frame difference gradients ( HDG ).

3.1 Optical Flow Directions
For each frame in the sequence we extracted the optical flow vectors using the im-
plementation [16] of the method proposed by [17]. From the f i =( x i ,y i ,vx i ,vy i )
optical flow vector its direction d i and magnitude m i is computed using
d i = 360
2 π

tan − 1 vy i
vx i

(1)

m i = vx 2

+ vy 2
i

, (2)

i

then the directions are linearly quantized. Fig. 1 (a) and (b) present two possible
arrangements. In our experiments we evaluated both types.

3.2 Frame Difference Gradients
Let g i =( gx i ,gy i ) denote the gradient of the absolute difference of two consecu-
tive video frames at position i ,calculatedbyusing3

3 vertical and horizontal
Sobel operator. The orientation and the magnitude of g i can be calculated by us-
ing Eq. 1. Finally, the orientations are quantized using an arrangement presented
in Fig. 1.

×

(a) (b) (c)
Fig. 1. (a) and (b) are two possible arrangements of quantization; (c) is the range of
directions (hue) in the HSL color space and is used for visualizing our features