As biometrics can now identify people based on individual measures, it appears prudent to translate these techniques for forensic use. As subjects can conceal features associated with identification, prior convictions have used gait and posture to identify suspect. The locations of human vertices are used within instantaneous posture matching. To derive a measure of confidence in this match, we use an automated analysis to determine the variation in the match measure as a function of increasing database size. We can match subjects between videos and assess the confidence in the match measure. We describe how we can derive a match for suspects recorded performing the same criminal act, in surveillance footage, and assess the confidence. As this is the first study of its kind, it raises many points to consider which can aid refinement not just of the matching procedure, but also constraints on the placement of cameras in surveillance.

We present a new method for view-point independent gait biometrics. The system relies on a single camera, does not require camera calibration and works with a wide range of camera-views. This is achieved by a formulation where the gait is self-calibrating. These properties make the proposed method particularly suitable for identification by gait, where the advantages of completely unobtrusiveness, remoteness and covertness of the biometric system preclude the availability of camera information and specific walking directions. Moreover, tests on the multi-view CASIA-B database, composed of more than 2270 video sequences with 65 different subjects walking freely along different walking directions have been performed. The obtained results show that human identification by gait can be achieved without any knowledge of internal or external camera parameters with a mean CCR of 73.6% using purely dynamic gait features. The performance of the proposed method is particularly encouraging for application in surveillance scenarios.

The simple command to use to convert a given video video.ext to a set of images preferably with the naming format XXXX.png is :

ffmpeg -i video.ext %4d.png

It will produce images with the naming : 0001.png 0002.png and so on. You can try different image format if needed ( Jpeg, bmp…)

In case you want to extract and convert only a specific portion of the video :

ffmpeg -ss hh:mm:ss -t hh:mm:ss -i video.ext %4d.png
-ss option is the starting point whilst -t option is the duration to be converted.

In recent years, automatic visual surveillance  has received considerable interest in the computer vision community. This is due to the increasing numbers of crimes from robbery to terrorist attacks, as well as the inability of human operators to monitor the increasingly growing numbers of surveillance cameras deployed in security sensitive areas such as government buildings and airports, or  public places such as shopping malls and streets. According to the British Security Industry Association, the number of CCTV cameras installed in the UK was estimated to be more than 4.25 million in 2004; this figure is expected to grow rapidly particularly after the terrorist attacks that London witnessed in July 2005.  Despite the huge increase of surveillance systems,  the question whether current surveillance systems work as a deterrent to crime is still debatable. Security systems should not only be able to predict when a crime is about to happen but, more importantly, they ought to identify the individuals suspected of committing crimes, say through the use of biometrics such as gait recognition.

Feature subset selection is the process of choosing the variables that are important for the classification stage from the original feature space. Feature selection is an important task for almost any pattern recognition problem (Webb, 1999). This procedure is aimed to derive as many discriminative cues as possible whilst removing the redundant and irrelevant information which may degrade the recognition rate. Furthermore, feature selection does not only reduce the cost of recognition by reducing the dimensionality of the feature space, but also offers an improved classification performance through a more stable and compact representation (Jain 1982). It is practically infeasible to run an exhaustive search for all the  possible combinations of features in order to obtain the optimal subset for recognition due to the high dimensionality of the feature space.  For this reason, it is recommended to use a feature selection algorithm as  the Adaptive Sequential Forward Floating Selection (ASFFS)  search algorithm (Pudil 1994).

The feature selection procedure fundamentally relies on an evaluation function that determines the usefulness of each feature in order to derive the ideal subset of features for the classification phase. For every feature or set of features generated by the feature selection algorithm, an evaluation criterion is called to measure the discriminative ability of the set of features to distinguish different subjects (Dash 1997). A number of methods (Mowbray, 2003) rely mainly on statistical metric measures which are based on the scatter or distribution of the training samples in the feature space such as the Bhattacharyya metric. These methods aim to find the features which minimize the overlap between the different classes as well as the inner-class scatter.

Since the adaptive background subtraction lacks capability to remove shadows, it is recommended using the approach described by Cucchiara (2003) to evaluate whether a foreground pixel corresponds to cast shadow based on the Hue Saturation Value (HSV) colour information. The chromaticity and luminosity of the foreground pixels are separated using the HSV colour space which is proved to match the human perception of colour more closely than the RGB model (Herodotou, 1998). The method proposed by Cucchiara et al assumes that shadows reduce surface brightness and saturation while maintaining chromaticity properties in the HSV colour space.

The Harris corner detector is a popular interest point detector due to its invariance to rotation, scale as well as illumination variation and robustness to image noise. The Harris corner detector is based on the local auto-correlation function of a signal; where the local auto-correlation function measures the local changes of the signal with patches shifted by a small amount in different directions.