When considering the application of CCTV in high-motion areas, such as street surveillance environments, a number of aspects are clear. Firstly, cameras are everywhere and placed anywhere. Secondly, the success of such systems is often touted to be high but ends up mediocre.
When these systems are evaluated, it appears that where systems are successful, this has often more to do with the control room operations than the technical systems themselves. In some control rooms the operation is efficient, suitable and trained staff is used. In most cases their diligence overcomes the shortcomings of technology.
Some factors are relevant with regards to the technology used. System components often fail and there is often neither the will nor the resources to maintain sophisticated systems. Secondly, the quality of technology is often not fit for purpose. According to the SAPS forensic department, more than 95% of video footage analysed by them turns out to be unusable for investigation purposes and even more are less useful for prosecution.
So what should be the criteria for a good street surveillance solution?
In order to define solutions, it is important to consider the critical aspects of street surveillance. (This discussion will be limited to CCTV and its options but other aspects of street surveillance are obviously also important). Street (CBD) surveillance consists fundamentally of:
* Detecting possible events, preferably pro-actively.
* Identifying real incidents and selecting actual incidents from false alarms.
* Managing these incidents remotely from a control room.
* Investigating incidents post-incident and gathering evidence.
* Presenting evidence to courts for prosecution.
What would be the characteristics of a system capable of delivering the above functionality and how can intelligent systems offer a contribution? In order to investigate what such systems can offer a brief definition of the characteristics of intelligent systems in this context would be in order.
Intelligent systems automate tasks, they operate without or with minimal human intervention, they can perform intelligent interpretation of events automatically and in context with the present scenario. Furthermore, intelligent systems can perform mundane tasks without supervision such as searching intelligently through video material for objects, incidents or even actions.
Let us briefly explore each of the above characteristics and the contribution of intelligent systems in each.
An important technical parameter when it comes to detection is the ability to recognise possible events. It therefore assumes that cameras are placed such that they cover areas where incidents are expected to happen and where the surveillance operator has jurisdiction to monitor. Clearly it would be a waste to simply place cameras everywhere. The cost of covering an area has reduced dramatically in the recent past but it is still costly when one considers installation, maintenance and so on. Covering an area is a subject where many strategies can be deployed to reduce cost, not the least sharing of cameras between multiple monitoring agencies – a possibility that has become significantly easier with the advent of digital cameras and networks.
Detection is often a question of recognising behaviour or anomalies. Thus, the appearance of objects (specific vehicles, specific people) or out-of-place behaviour such as movement in dead spots, running, etc, may trigger investigation. Humans are particularly good at these tasks but suffer from human frailty (lack of alertness, limited concentration and others). When it comes to camera coverage for this purpose, broad area coverage is required, but sufficient resolution to perform adequate detection is also required. While behavioural analysis does not necessarily require high detail, as motion in large spaces are analysed, specific recognition such as facial or ANPR does require much better resolution.
Wide area detection sensing will therefore probably consist of wide-angle views to cover most of the area with specific high-resolution (narrow angle) views on specific places monitored for specific events. With wide area sensing illumination becomes a real problem as cameras have notoriously low dynamic ranges and special care must be taken to offer constant level illumination or support illumination.
What can intelligent systems offer?
In the detection arena, intelligent systems strive to improve detection and, in particular, automated detection. This will assist human operators by taking over the mundane and boring tasks of continuous monitoring. But to be useful it must be able to offer some capabilities such as continuous detection, improved detection, filtering true alarms from false/nuisance alarms and so on.
Modern image processing techniques can support this function successfully by offering both behaviour detection (motion and tracking analysis) as well as specific object detection (facial, ANPR, shape recognition). In addition, analysis of detected events can, by using sophisticated analysis and cross-correlation techniques, significantly reduce false alarms. Analysis of historic events and time-line analysis can create pro-active trends and contextual analysis can be used to create priorities in detected events.
Behavioural analysis is a hot topic today and, while still in its infancy, it can be used to create tracks of movement and predict destinations, time to reach possible targets and, even to some extent, intent of the action. These do not have to be from single views but can be achieved in a combination of multiple camera views, combining information into single events. Such analysis in conjunction with scenario analysis can offer very powerful incident detection capabilities.
Last but not least, because machines can work in parallel and process a lot of information simultaneously, higher resolution cameras can be used with no loss of information. (When cameras with resolution higher than the screen resolution are used, some pixels are simply not displayed and no one can observe what happens unless you zoom in – an action that will only be performed when something has been detected in the first place.) This means that fewer cameras can be used to achieve the same result, resulting in significant cost saving.
All of this serves to improve the quality of service, assist humans and reduce reaction time in addition to potentially lower the cost of providing security and indeed even of hardware required.
When it comes to identification, the requirements change dramatically. Here identification becomes paramount and therefore resolution is critical. Fortunately, the use of pan/tilt/zoom cameras have also become much cheaper and their capabilities have also increased dramatically. These cameras are very poor in their ability to monitor as they have essentially tunnel vision and have to be directed if used in the monitoring sense, providing at best, a time sliced view of any area. But when it comes to identification they can zoom in to any object and because of their narrow view the camera is essentially subject to a very even illumination spread. With the use of low-light capabilities, or support illumination, very good identification can be achieved.
In this context, humans offer at present by far the best abilities, so what can intelligent systems offer? A number of automated tools can be used to assist humans in this detection. These include rapid pointing of the camera by pointing at a position on the static (wide area) scene and automatically designate the camera to slew to this point. Similarly, immediate facial recognition or number plate recognition can identify objects specifically assisting an operator in knowing who he is dealing with based on prior information. The automatic identification of objects allows operators to rapidly classify an event into an incident and also to prioritise incidents according to historical information that may have been gathered. This ability to bring scenario analysis into decision making is an area that promises huge advances in security provision and particularly in high dynamic environments such as street surveillance control rooms.
Up to this point, machines could offer a significant contribution to the surveillance process. However, when it comes to incident management, humans are in a class of their own. Incident management includes a number of functions such as continuously monitoring the event itself, predicting possible scenarios and monitoring these potential spots, understanding what resources are available to manage the incident, communicating with these resources and so on. This is clearly a big job and is often handled by an incident management team.
The task requires amongst others the following of the command and control system:
* Presentation of visible views of all views that are relevant to the incident and removal of irrelevant views.
* Understanding the environment and location by presenting the team with maps/building layouts of the relevant areas.
* Offering additional information, such as other events that are happening in the area that may be affected by, or influence, the incident.
* Be in constant communication with resources and offering them useful information as well.
While this is clearly a human task, intelligent systems can greatly assist in the following way:
* Enhancing the views that operators have of the incident. This could include a variety of image enhancement to allow for detection of weapons, look into dark spots, identifying people, controlling the camera to search for subjects, etc.
* In addition, immediate access to relevant recordings of the incident, as well as access to historical information such as names, addresses and prior suspicions, also assist in contextualising the incident and managing it appropriately.
* The presentation of relevant views can be automatic by the use of scenario planning and analysis. Such plans could have foreseen certain disasters and initial war room views could be pre-set to speed up reaction.
Once the excitement has died down the real work often starts. Whether the incident was contained or not the analysis of what led up to the event and what happened during the event must now take place. This requires access to historical information and this is also where many systems fail. It is probably safe to say that cameras and video recording equipment are mostly sold on their ability to compress video streams into narrow bandwidths, or their low storage requirement resulting in long storage periods.
It is sad to say that in most cases the actual use of the information is disregarded to save money. The design should consider what the information is to be used for and once this is defined, the best, most cost effective solution that will still achieve this should be acquired. Consideration should be given to video resolution, camera dynamic range, required frame rate, storage period and so on and at the risk of stating the obvious some aspects should be re-stated.
Video resolution – this must be such that the type of object is recognisable at the distance seen. Thus, to recognise that an object is a car one would need at least three pixels per metre on the object. To recognise the shape and model at least 10 pixels per metre, to identify its number plate at least 500 pixels per metre. Similar specifications for identifying all objects must be calculated and appropriate camera specifications be used.
Dynamic range – this is difficult to define as it is seldom specified by manufacturers, but if objects are to be identified over the image a contrast difference between subject and background of at least 0,5 dB (for simple processes) is required throughout the scene, ie, in shadows as well as in bright areas.
Frame rate – a parameter which is often overstated to the detriment of resolution. Frame rate is only required to be high if the possibility of missing an action exists. For most views frame rates of less than 8 fps is adequate to determine behaviour and if less of an eye strain is important due to long viewing periods, frame rates of 10 fps to 12 fps are usually sufficient. Whenever identification is important resolution should not be sacrificed.
Storage period – the general rule of thumb is that the storage period need not be longer than the period required to recognise the crime. Once this is done relevant information can be extracted and saved into long-term storage.
Given that the above was achieved, how would intelligent systems assist at the post-event analysis level. One of the most obvious tasks would be to search through available historical events of both the incident and the pre-incident information. Intelligent machines can locate sightings of people, vehicles and other events prior to the incident and it can do this automatically. Using track analysis, combined analysis from multiple sources and so on they can assist in powerful analysis of information creating very good evidence for prosecution. An intelligent system should be able to treat recorded video as live video and perform all the same detection and recognition functions on historical information – often faster than real-time.
Evidence for prosecution
While the prosecution effort is generally not part of the command and control system and at this point, the use of intelligence is somewhat limited to that of the lawyers and advocates. It is, however, important that evidence is managed in a way that is conducive to successful prosecution. This requires amongst others:
* Evidence be packaged in formats that prosecutors can easily access and present.
* Evidence is treated in such a way that it cannot be tampered with or be lost.
* Evidence can be stored securely until the trial is complete.
Most of these functions, other than packaging of the evidence and the way it is packaged, are primarily the responsibility of the prosecution agency.
It should be clear that the next wave in security systems and in particular, high-motion systems will be the use of intelligent systems. These have the potential of reducing human workload, but more than that, it has the potential of using humans in a way where their unique and powerful abilities are best used, resulting in improved security at a lower cost.
Dr Coetzer can be contacted at Protoclea Advanced Image Engineering ( email@example.com or www.facebook.com/protoclea), a company specialising in the development of products such as those described in this article.
|Tel:||+27 11 100 0935|
|Articles:||More information and articles about Protoclea Advanced Image Engineering|
© Technews Publishing (Pty) Ltd | All Rights Reserved