July 11, 2022
Video analytics can provide an intelligent and automated approach to perimeter security by detecting unauthorized intrusion along any perimeter in the lowest light conditions and on complex sites. In the following article, we discuss the challenges faced by video analytics and how we overcame these challenges based on our experience and customer feedback.
Historically, with analog cameras, you would have had a security guard watching a video wall and looking for events to happen. The downside of this was user fatigue; a person cannot watch everything all the time, resulting in missed events.
With the rise of IP cameras and video analytics came the ability to automate video stream analysis and detect objects, such as people or vehicles moving, within the scene. When an event is detected, an alarm will be triggered and sent across a network to software that can automate the response. The event alert gives the ability to investigate an incident faster.
Video Motion Detection (VMD) relates to analytics for detecting motion on a video. VMD does not detect objects specifically but frame-by-frame changes or movement on the pixels. The advantage is that you could start recording only when pixel changes are detected, significantly decreasing the bandwidth and storage required for video recordings, resulting in an easier event search.
The environment and lighting can challenge VMD, and false alerts can happen with changes in foliage, shadows, branches, animals, etc.
However, it is also possible to configure filters to improve detection and reduce false alerts. Filters may include minimum and maximum object height.
It was written in IPVM that “Integrators reported it was easiest to use cameras’ built-in analytics instead of using add-on or third-party analytics.”
There is much discussion in the industry about the advantages of edge analytics versus a server or the cloud. While cloud analytics has various advantages, edge analytics can help maximize legitimate detections of intruders.
If the server or cloud does not have pre-processing performed on the edge, they won’t scale well, and data loss due to video compression will affect the efficacy of the analytics detections.
In addition, server or cloud-based solutions may suffer from latency and bandwidth limits or restrictions.
Edge analytics provides the ability to process what is happening in a field of view and discern if a relevant alert is triggered, ensuring faster detections and less expensive infrastructure. Edge analytics run on raw video data instead of the post-processed encoded video on a server, allowing the analytics to gather more sensitive and accurate data.
Thermal cameras have several advantages over visual cameras regarding edge analytics. Thermal cameras do not need to rely on VMD alone; a thermal camera detects heat and uses heat to build an image. Therefore, instead of motion alone, thermal cameras have the additional benefit of utilizing the heat of an object and its movement to determine whether it is a valid target for analytics.
Thermal cameras do not need to rely on light and, therefore, can detect targets 24/7 regardless of lighting or inclement weather.
When performing analytics, we use the uncompressed 16-bit sensor data from the camera, with a maximum of 65,536 tonal grey values, to ensure the highest detection capabilities. Once compressed, the image has 8-bit visual data (H.264 or MJPEG output), and a maximum of only 256 tonal grey levels is achievable, which stretched across the scene’s dynamic range, will reduce the ability to detect a target with server-based analytics. The number of grey levels is less relevant to visual detection with the eye but crucial for accurate analytics.
Noise Equivalent Temperature Difference (NETD), or sensitivity, plays a big part in the camera’s ability to detect movement effectively. Most camera manufacturers will advertise a NETD of around 50°mK, which means that the sensor has a ΔT (Delta T), or a minimum difference of temperature between two measuring points of 0.05°C; however, you would also need a lens with an f-number of 1.0 and an ambient temperature of around 25°C to achieve this. In addition, this would apply to the 16-bit uncompressed sensor data.
When compressed, the image decreases to 8-bit, and the sensitivity will drop between 200 and 500°mK (dependent on multiple factors), which means the minimum temperature differences between two points in the scene can be as high as 0.5°C, a difference of 0.45°C when compared with a 16-bit image. How does this affect detection?
Temperature difference is an effective parameter for analytics detection, and the minimum temperature difference is a vital factor. If a person in the scene is against a background measuring 25-30°C, visually, it may be hard to distinguish between the grey levels. However, suppose the minimum temperature difference of a 16-bit uncompressed image is 0.05°C, compared to 0.5°C with an 8-bit compressed image. In that case, the camera’s analytics will have a greater chance of detecting that person against the background using the edge analytics on the camera.
The edge analytics perform the analysis based on Regions of Interest (ROI) and use a user-defined minimum and maximum size of an object. The analytics will consider each pixel and, based on the minimum and maximum dimensions and assuming a constant slope, will decipher the scene by determining a range for each pixel. By setting the maximum speed of an object, the camera can accurately calculate its speed and based on size and movement, can determine the object’s relevance, e.g., if it’s too fast and too small, such as a bird, it will be discarded.
The analytics has a scene learning feature to learn the background scene in the camera’s field of view.
For the learning to be effective, the scene should be clear of moving targets (people and vehicles) and learn the scene over several hours, including at difficult times, such as dawn and dusk.
The scene learning will analyze the movement of trees and vegetation against the background (sky, road, or open area) relative to the temperature differences.
Using the knowledge camera has from the object sizes, maximum speed, and the information gleaned from its learning process, it can effectively classify and detect a human, animal, or vehicle with high accuracy. The camera requires less than 6 pixels for detection and is effective in all weather conditions due to analysis performed on the 16-bit sensor data.
Suppose a target moves behind trees or vegetation, and the camera has previously detected this specific target. In that case, the camera will continue to track the pixels of the target that continue to meet the requirements of detection.
AI or Artificial Intelligence is a buzz in the industry, but VMD continues to dominate as it is simpler and more efficient. In contrast, AI is still in its infancy, requires more power, more initial setup, and is more expensive.
We are looking at adding AI functionality to our cameras, which will undoubtedly be the future of video analytics. However, as the technology is still in its infancy and less cost-effective than current solutions, good VMD edge analytics will retain a competitive advantage in the foreseeable future.
When buying a thermal camera, it is essential to know that a camera with 35°mK sensitivity is not necessarily better than a camera with 70°mK sensitivity; ask the camera manufacturer whether the edge analytics are processed pre- or post-video compression.
If the camera does not have edge analytics but relies on third-party server-based analytics, a camera that declares 35°mK sensitivity will have a sensitivity somewhere between 200-500°mK after compression and considering variations in environmental conditions.
The additional downside of server-based analytics is a compressed image and less sensitive analytics detection, potentially with more false alarms. Also, a server-based solution will require greater bandwidth and infrastructure costs.
July 11, 2022