How to Use Intelligent Video Analysis to Improve Search Efficiency

The application bottleneck of traditional mass storage technology For a large-scale video surveillance system such as Safe City, there are massive amounts of video stored in the system every moment. These videos are crucial for fighting crimes and safeguarding people’s lives and property. Currently, a large-scale video surveillance platform based on a distributed system architecture design adopts a divide-and-conquer method and integrates various strategies such as front-end storage (NVR/DVR) and back-end storage (storage server, magnetic array, etc.). From a technical point of view, the storage problem of massive video can already be solved more steadily. However, the mainstream search methods based on time points and preset events are too simple compared to the current complex security situation. The following takes an actual search requirement as an example to analyze the traditional search methods.

Retrieving needs: Searching for men from 10:00 to 12:00, who appear near the main entrance of XX Bank and wear white shirts.

In order to find a video that meets the requirements, a lot of manpower will be required for manual analysis in a traditional video storage system. The traditional monitoring system can accurately find the cameras near the main entrance of XX Bank through GIS, find the relevant video through time retrieval, and then need to rely on human to conduct manual viewing. Even under 16x speed high-speed browsing, further finding related videos is also a It is a waste of manpower and time, and due to people's participation, long-term viewing leads to a high risk of missed inspections.

We can see that the traditional search method is awkward when it comes to fulfilling this search requirement. The search method is not "smart", which has become a bottleneck in the application of traditional large-scale monitoring systems.

Taking into account the complexity of specific cases, today's request for "white shirts" may become "flower shirts" tomorrow, so users not only require that searches be more "smart", but also adequately "flexible." Of course, the user also wants the system to complete the query as soon as possible. Then, we can sum up the user's retrieval requirements for massive video storage systems to be smarter, more flexible, and faster.

Using intelligent analysis techniques to improve the effectiveness of search algorithms Introduce intelligent video analysis technology to achieve "smart" and "flexible"

Since users urgently need to search more "smart", it is imperative to use intelligent analysis technology to improve the system's "intelligence."

The current intelligent video analytics technology applied to video surveillance can be summarized as three mutually dependent technical pedigrees: target extraction technology, target classification (identification) technology, and rule judgment technology. These three types of technologies constitute the three basic modules of the intelligent video analysis system: target extraction module, classification identification module, and rule judgment module.

First, the digital video stream is video-decoded and decoded into a video or image sequence that the intelligent video analytics system can recognize as an input to the intelligent video analytics system. The analysis system first uses the target to extract the input image sequence to perform pre-background separation to complete the target extraction. The more complex intelligent video analysis system generally has a classification function to identify the identified target and identify more attributes of the target. For example, the more common classifiers include: people car classification, color classification, behavior classification, etc.; then the system will be extracted the target and its classification information, to the rule judgment module, rule inference, if the violation of certain settings Rules, the system will output alarms.

Putting the search requirements mentioned above into the above process, we can find that the existing intelligent video analysis technology can theoretically help fulfill this demand. "White shirts, men" These are the identification requirements for the classification identification module, and also the rules in the judgment inference module. In other words, the so-called "smarter" corresponding intelligent video analysis technology requires the classification and identification module to be more powerful; the demand for retrieval must be constantly changing, for example, "white shirt" becomes "flower skirt" and "man" becomes The "woman" and "flexible" requirements require the rule judgment module to complete.

When the user initiates a query, the rule generator is first used to translate the query request into rules and enter the rule inference module. After that, the system performs intelligent analysis on related videos, that is, target extraction and target classification and identification, and rule inference judgment and identification. Meet the requirements of the video.

This model, with the help of intelligent video analysis technology, can improve the “IQ” of the query system and achieve users’ demands for “smart” and “flexible”, but how effective is this system? There is also a question to be considered - whether it is possible to accurately estimate the exactness of the target extraction and model matching for the current intelligent video analysis technology, so as to judge the "men and women" and distinguish between "white shirt" and "white windbreaker" different. These problems are difficult to solve in the existing analysis techniques and algorithm models. In summary, the problems with this model are as follows:

First, the efficiency is low, and the query speed depends on the analysis speed.

The second is that for the actual requirements in practical applications, the situation that the intelligent technology cannot be achieved has not been dealt with.

The reason for using the video analytics technology to construct the “cue” database model 1 is that the reason for its low efficiency is mainly due to the need to intelligently analyze possible goals and the intelligent analysis algorithm takes time to complete. This process limits the efficiency of this model.

In fact, for this problem, we can increase the search by adding real-time video intelligent analysis server to intelligently analyze real-time video while storing, and save the results of the target extraction module and classification identification module to the database. s efficiency. This idea is similar to the video retrieval technology used by Google. It is just different from the video source of the extraction target and the classification analysis, and the timing for preprocessing is different.

Of course, this preprocessing process based on intelligent video can also be completed by the terminal device. As for how the target device extracts the target module and the analysis result of the classification and identification module is sent to the video monitoring platform, reference can be made to the relevant part of the OnVif protocol.

When querying model 2, firstly, it converts the query conditions, converts the query conditions into rules that the system can support, generates query statements matching the database according to the rules, and then finds related videos according to the query results.

About "misunderstanding" and "missing check"

Contrary to the current state of the art of intelligent video analytics, at least the shirts and the men are difficult to distinguish among the search requirements of "a man wearing a white shirt at 10:00 to 12:00, near the main entrance of the XX Bank." Taking into account the light and other factors, the resolution of white is also difficult. That is to say, even if the intelligent analysis technology is used, due to the limitation of its classification and identification module, there are also misunderstandings and missed detection possibilities for some inquiry requirements. For practical applications, the risk of missed detection is greater than the error. Based on this, it is necessary to introduce some methods based on probability theory, such as Bayesian method, to alleviate the pressure of missed detection.

Conclusion In summary, the introduction of intelligent video analysis technology can greatly improve the retrieval efficiency and hit rate of the original massive surveillance video storage system. However, due to the current status of intelligent video analysis technology, there are still some technical risks in this solution. Although some probabilistic methods can be used to introduce concepts similar to "similarity" to mitigate the risk of "missing", this scheme is still a long way from the final user requirements. However, from another perspective, distance produces beauty. The reason why technology is so attractive is largely because of the existence of "distance." There is reason to believe that, with the rapid development of the industry, the introduction of intelligent analysis technology in mass surveillance video storage systems will become a trend.