Leandro da Cunha, Surveillance BU Executive at Duxbury Networking
A surveillance system can store days or months of footage from multiple cameras. When an incident occurs, the relevant evidence may occupy only a few minutes in that archive. Finding those minutes has traditionally depended on knowing where and when to look. Investigators often begin with a time and camera, then work through the footage until the relevant sequence appears. New AI tools are beginning to reduce how much of that footage a person has to review manually.
One such example is Milestone’s XProtect Video Summarization. It can analyse selected clips and generate natural-language descriptions of what is happening in the scene. The organisation has also announced AI Search, due by the end of this year, which will allow operators to describe what they are looking for in ordinary language and surface relevant video snippets.
It gives operators another way into the archive when they do not yet know exactly where to look.
Searching without the exact timestamp
Video search becomes difficult when the investigator knows what happened but not exactly when or where it happened. A description, such as a vehicle entering an area, is much closer to how somebody remembers an incident than a timestamp. AI Search is designed around that difference. The operator describes the event in natural language, and the system looks for relevant footage.
XProtect Video Summarization comes into play once there is footage to examine. The plug-in uses a vision-language model to analyse selected video and produce a description of the activity it detects. That response can be saved as a bookmark description in XProtect and searched later using plain text.
That can reduce the amount of footage an operator has to work through before finding the part they need.
The time saving is in the review
The volume of stored footage keeps growing as camera estates expand and retention periods become longer. That leaves investigators with more material to work through when something happens.
Reviewing footage and documenting an incident still involve a lot of manual work. Video Summarization can run on demand or be triggered by events in XProtect, giving operators a structured description of selected footage without requiring them to watch the whole clip first.
If an incident spans 40 minutes of recordings but only a short sequence is relevant, a useful summary can help the operator decide where to begin a closer review. It can also give the investigation team a more consistent starting point for documenting what happened.
The recording still needs to be checked
Generative AI can produce descriptions that sound convincing even when they are not completely accurate.
Milestone addresses this directly. Its guidance says hallucinations are reduced by grounding responses in actual video frames and detected objects. It also says generated results should be interpreted in the context of the underlying video evidence.
A summary can point an investigator towards relevant footage and help explain what appears to be happening. The operator still needs to be able to return to the recording and review it before reaching a conclusion.
Privacy stays part of the design
Video is sensitive data, so organisations also need to understand how footage is processed when these AI tools are used. XProtect Video Summarization sends selected clips securely to its Hafnia cloud instances in the US or EU for analysis, with submitted video destroyed immediately after inference. It is also developing on-premises and air-gapped options for environments that require local processing.
Customers need a clear view of where selected clips are processed and who is allowed to use the tools. Those decisions also need to fit the organisation’s existing evidence procedures.
What this changes for investigations
For Duxbury partners, the focus is on how these tools can make it easier for operators to work with their existing footage.
The value becomes clearer in environments with large camera estates or long retention periods, where finding the right sequence can consume a great deal of an investigator’s time. Natural-language search and summarisation can reduce that effort without removing the operator from the investigation.
Duxbury works with partners on Milestone XProtect and the wider surveillance environment, helping customers design systems around how video will be used after an incident. Within that environment, XProtect provides the VMS layer connecting surveillance devices and analytics within a common management environment. Surveillance systems already capture more footage than any operator could realistically review in full. Making that evidence easier to find may prove just as valuable as capturing more of it.
