Two Ways to Evaluate Intelligent Video Analytics Surveillance Technology

Not all products with intelligent video analysis functions can be classified as "intelligent". The determination of "intelligent" products depends on intelligent evaluation - by observing the function or size of its intelligent video analysis function, etc., it can help users choose video surveillance system equipment products with the required level of intelligence.

One of the evaluation methods is to see what level this intelligent product belongs to. By evaluating the level range of the product, we can understand the role of the intelligent video analysis software algorithm, and basically solve the problem of the level of intelligence of the video surveillance system. In this way, users can choose according to their own security needs to ensure the maximum security benefits. For specific evaluation methods, please refer to the article "On the Evaluation Standards and Methods of Intelligent Video Surveillance Systems". That is, the role of intelligent video analysis is evaluated by "the level division of the intelligent evaluation method with 5 levels of scoring". For example, the products in the first level range have very poor intelligent video analysis and are not practical at all; the products in the second level range can identify people and objects (i.e. the size or shape of the target), but cannot judge the abnormal behavior of people and objects, and their intelligent video analysis is not up to standard, so the products cannot be practical; the products in the third level range can identify certain abnormal behaviors of people and objects under indoor conditions such as room temperature, and their intelligent video analysis is OK, that is, it reaches the standard, so the products can only be limited to applications under indoor conditions such as normal temperature; the products in the fourth level range can detect abandoned objects (discarded objects) in crowded scenes, and can very effectively detect luggage left alone at airports, or items that may be stolen in museums and warehouses. Generally, they can be linked with anti-theft, fire prevention, access control and other systems, and can accurately detect and identify people's faces, gait or voices. Products or systems in this level range can identify abnormal behaviors of people and objects under general outdoor conditions. Their intelligent video analysis works well and their anti-interference ability is good, that is, they are good, so the products can be used indoors and outdoors. Products in the fifth level range can detect abnormal behaviors of people and objects in the above levels under more severe climatic conditions than those in the fourth level range, and can detect more comprehensive and more abnormal behaviors of people and objects, such as people wearing masks, pointing knives or guns at people, etc., and can accurately link with anti-theft, fire prevention, access control and other systems, and can accurately detect and identify people's faces, gaits, voices, etc. Because it can eliminate all interference and can be used in extremely harsh environments, this level is excellent, and its intelligent video analysis works very well, that is, the level of intelligence is extremely high (this is the direction we are working towards).

The second evaluation method is to evaluate the role of intelligent video analysis and recognition technology from the perspective of intelligent software algorithms. Because the core of intelligent video analysis and recognition technology is intelligent software algorithms (see my article "Intelligent Software is the Core of Intelligent Video Surveillance System" published in 2008). Therefore, we can start from the perspective of intelligent software algorithms to evaluate whether the role of intelligent video analysis and recognition technology is truly applicable. Generally speaking, effective intelligent video analysis and recognition technology, that is, practical intelligent software algorithms, mainly depends on whether they can meet the following requirements.

1. It depends on the degree of consistency between different monitoring scenarios and the mathematical model of the intelligent software algorithm. Whether the mathematical model of the intelligent software algorithm used takes into account the influence of factors in different monitoring scenarios, that is, whether the monitoring environment and the intelligent software algorithm model achieve the greatest consistency. For example, when the target size is used as an important classification feature, usually in a scene with a large depth of field, the target size varies greatly, and the accuracy of target classification will be greatly reduced. This can be solved by reducing the depth of field of the scene, or adding a scene calibration algorithm, or reducing the weight of the size feature in the classification algorithm.

2. Check whether some software modules are added to the intelligent software algorithm framework according to different scenarios:

For video images captured by cameras mounted on moving objects such as cars, an anti-shake software module must be added to improve the processing effect in the case of camera shake;

In outdoor scenes with severe shadows,

It is necessary to add a shadow suppression software module to improve the processing effect in outdoor scenes with severe shadows;
In scenes with drastic changes in lighting, it is necessary to add a light change suppression software module to improve the processing effect in scenes with drastic changes in lighting;

In scenarios where target images frequently occlude each other, an occlusion processing software module can be added to improve tracking accuracy in scenarios where target images frequently occlude each other.

3. Whether the impact of complex surveillance scenes such as low illumination, high disturbance, and high crowding on system performance is taken into account. For example, whether there is a tidal filter, that is, filtering out objects that often change in shape or move in an irregular direction or too fast, such as sunlight reflected from the water surface.

4. Are there object size filters (i.e. filtering out objects that are too large or too small) and object size sudden change filters (i.e. filtering out objects whose size changes suddenly) to avoid false alarms, etc.