Video investigations
From hours of footage to a focused review: using natural-language video search
By Rong Kai Wong
How property and security teams can structure video searches, connect observations across cameras, and validate results without treating AI matches as proof.
Turn an incident report into observable details
A request such as 'find the incident from last night' leaves a reviewer with little to work with. Start instead with visible details: an object, an activity, a clothing description, a vehicle, a location, and a time window. Avoid adding assumptions about identity or intent that the footage cannot establish.
Natural-language video search lets a team express those clues in everyday language. SemanticIQ is LISE's video-record search and facial recognition capability. Its role is to narrow the material for review, not to replace the investigation or guarantee that every relevant moment will be found.
Search broadly, then refine one clue at a time
Begin with the clearest observable feature and the most credible time range. Review the returned footage, then adjust the wording or location if the results are too broad. Changing one part of a query at a time makes it easier to understand why the results changed.
Keep a record of the query, the cameras searched, and the time range. If the investigation expands later, another reviewer should be able to understand what was checked and which parts of the archive remain unreviewed.
Build a timeline across camera views
An appearance on one camera is only one observation. Review adjacent views and relevant recording periods to establish a sequence. Account for camera-clock differences, gaps in coverage, and changes in lighting before connecting two clips.
Similar clothing or a similar-looking vehicle does not establish that two observations show the same person or object. Treat a suggested visual match as a lead to examine, especially when the available detail is limited.
Use identity-related features with additional care
LISE's SemanticIQ can use facial recognition alongside visual cues such as clothing and colour to recognise and tag a person of interest, helping reviewers trace that person's appearances across the connected camera network. The software does not know the person's actual identity: a person-of-interest tag links relevant footage, rather than identifying someone by name.
If a client associates a tag with an actual identity, the client is responsible for keeping that identity information confidential and secure. Footage and face-based matching still require appropriate access controls, retention practices, and privacy safeguards even when names are not recorded. Image quality, angle, and occlusion can affect matches, so reviewers should corroborate results rather than treat a suggested match as conclusive proof.
Measure the quality of the review, not just speed
Compare time to a useful clip with the existing review process, but also examine whether important footage was missed. Test known scenarios and include a manual check of representative recordings to understand the system's limits.
The useful outcome is a better-supported investigation with an understandable record of what was reviewed. Faster search is valuable when the team can still explain its conclusions and preserve relevant evidence under its own procedures.