Results
Results of HOG+MS on Sequence#3:
Results of HOG+MS on Sequence#4:
Conclusions
In this project was presented a new approach to capture images in indoor sports venues which will
be used extract valuable information about collective performance. Unmanned Air Vehicles as
Quadcopters allow a cheap, portable, flexible and reliable platform to acquire images from indoor
sports events, in the particular case of this research, indoor soccer. However, due to their own
dynamics and also external factors is impossible to avoid drone’s undesired motion which will
result in a series of problems not typically found neither in the literature nor in the commercial
solutions available on the market. In this work was set a framework to extract useful and reliable
information from indoor soccer games composed by different stages:
Video Stabilization was required to maintain spatial coherence of pixels intensities despite
the drone’s motion. Using FAST features and RANSAC matching between adjacent frames
is possible to estimate the inter-frame motion and consequently compensate it. The method
proposed relied on its simplicity and efficiency on the test sequences. It can deal with the
high frequency jittering of the camera but over time error is being accumulated and not all
the movement is compensated. If longer sequences were tested or if it was intended camera
motion to cover all the action of the game, more complex methods for stabilization would
be required.
Camera calibration is an essential stage of computer vision systems, in this project it is
essential to map the position of the players in the field from their coordinates on the image.
Since camera movement is not totally compensated is necessary an automatic and dynamic
method for calibration. In this project is proposed a simple method based on detection of the
lines marked on indoor sports venues and the posterior match with the lines of the virtual
model created manually. The results proved that calibration does not drift. Correction rate
is an important parameter to be set.If it is high, it will demand huge computational power
and if it is low lines matching can fail and consequently the calibration too
Since most of the common methods to player detection are not suitable to this project due to
nature of the image acquisition system it is proposed a methodology based on HOG people
detector with short term position estimation with mean shift tracking. The detection is based on HOG descriptor and a classifier trained with a dataset of people on upright position. This
detector has low precision to detect players in sports scenes. False positive handling and
team identification was carried using histogram comparison in the RGB colormap and with
a classifier based on k-Nearest Neighbour. This is presented as an simple and robust solution
for histogram comparison but not efficient computationally. After this stage precision increased
notoriously, on the other hand recall decreased on the same proportion. To estimate
players’ position while HOG detections are not available it is performed mean shift tracking
which will find on the posterior frame the location of the image that maximizes the similarity
with the current appearance. Final results presented a precision and recall rounding the
75%. The algorithm shown difficulties to deal with players entering and leaving the image
since is there non prediction of where and when a new track must be created. This was the
main cause to precision decrease from the HOG detection with false positive handling.
Finally some methods are proposed to extract high level information from the data corresponding
to players’ positions on the field. These methods were based on the common
knowledge of the authors about the game and evaluated considering only subjective criteria.
Even with a not totally precise low-level information it was possible to infer some high level
interpretation related to field occupancy, offensive trends and defensive tactics.
Considering the final results presented is possible to assume that the proposed goals for this
work were achieved despite most of the methods presented can be upgraded and refined to achieve
most accurate results mainly on players’ detection and tracking. In order to extract robust and truly
useful information, drone’s camera has to be able to cover the entire field. This can be achieved
using multiple drones or with an automatic flight control such that it could follow game action
based on ball or players’ position.