Image Processing systems

Image processing systems (IMS) generally have better accuracy compared to the EMS, and an improved range when compared to the OMS. In image processing captured films or photos are digitally analyzed. Oppositely to the other measurement methods which are sensor-based, this method is vision-based, using optical cameras and computer vision algorithms. This marker-less tracking can be a big advantage in sports, such as for event-detection (Zhong & Chang, 2004). Image processing also has some drawbacks: it is not easy to perform image recognition in real-time and it might require expensive high quality and/or high speed cameras. The accuracy is also dependent on the experimental set-up, namely the position of the camera in relation to the object trajectory, and the number of cameras (Lluna, Santiago, Defez, Dunai, & Peris-Fajarnes, 2011). Furthermore, generally, an increase in camera resolution results in a decrease in feasible maximum sampling frequencies.
Vision based systems can be divided into two categories: Model-based tracking and feature-based tracking. Model-based tracking uses a 3D model of the tracked object. In the basic concept of the model-based tracking, the pose information is updated in each video frame, first by using a dynamic model via a prediction filter and then by measurements in the video frame. A drawback of model-based tracking systems is that they are hard to use in unknown environments and restrict camera motion, due to the necessity of additional information such as 3D models of participants and environment (Bader, 2011; Ceseracciu et al., 2011).

Feature-based tracking algorithms use interest points in the frames to track the object. There are two kind of feature-based tracking algorithms: marker tracking, which uses known-markers, and marker-less tracking, which focuses on tracking 2D features such as corners, edges or texture (Akman, 2012). Note that the marker tracking in IMS differs from OMS, because IMS uses (for humans) visible light, whereas OMS works with infrared light.

For marker tracking, known-markers are used to track the object. This is usually more accurate than to detect natural features (e.g. existing corners or edges), however the markers must be put precisely in place before the experiment (grid set-up) and occlusion of markers may occur. In sports, marker-based feature tracking has been applied in the collection of kinematic data on a ski and snowboard track, where an accuracy of 0.04m was obtained in a 2500 m2 range (Klous, Müller, & Schwameder, 2010).
Marker-less tracking eliminates the dependency on prior knowledge about the environment and extents the operation range. This natural tracking is a hot topic in, for instance, robot vision and augmented reality. However, in those applications, the cameras are actually attached to the object that is being tracked, in contrast to the sports application, where, up-to-know, the camera is static, while panning, tilting, and/or zooming (Liu, Tang, Cheng, Huang, & Liu, 2009). Liu et al. (2009) mounted a panning camera to the ceiling to track short-track speed skaters during a match, using a color-histogram of the skaters; they obtained an accuracy of 0.23 m (area 810 m2).
The KinectTM sensor – which was originally designed to allow users to interact with a gaming system without the need of a traditional handheld controller – can also be classified as a marker-less tracking device, although the working principle is slightly different from what was previously described. The system projects an infrared laser speckle pattern onto the viewing area of the infrared camera. This infrared camera detects the pattern and enables the creation of a 3-D map by measuring deformations in the reference speckle pattern. Due to its low-costs and reasonable accuracy (0.19 m at 7.5 m2 (Dutta, 2012)), the device is often used in scientific research (Bonnechere et al., 2014; Choppin, Lane, & Wheat, 2014; Dutta, 2012). The drawback of the Kinect camera is the small field of view; furthermore, the system struggles with the detection of dark surfaces that absorb light, shiny surfaces that result in specular reflection and rough surfaces if the angle of incidence of incoming light is too large (Dutta, 2012).

At present, available computer-vision-based measurement systems are outperformed by either optoelectronic or electromagnetic measurement systems and their maximal range is small. Although no mature system exists at the present (July 2017), a large number of open source codes are available and progress is rapid (Scaramuzza & Fraundorfer, 2011). Open-source databases with human kinematic data are provided to enable developers to verify their algorithms (HumanEva, 2017). This not only enables the verification of the developed systems, but also eases the comparison between systems for researchers developing their study setup.