Video today is passive.

You scroll. You scrub. You guess timestamps.

SceneIQ changes that.

It transforms video into queryable structured intelligence — not frame detection, not static labeling, but temporal reasoning over evolving events.

User types:

"Football World Cup trophy celebration"

Within seconds, SceneIQ:

1. Detects trophy object
2. Identifies crowd motion spike
3. Tracks multi-player clustering
4. Detects raised-arm celebration pattern
5. Computes interaction density
6. Ranks scene importance

Returns:

Scene:               Trophy Lift Celebration
Timestamp:           01:42:18 – 01:44:03
Importance Score:    0.93
Motion Intensity:    High
Interaction Density: High
Detected Objects:    trophy, players

Video jumps instantly to the exact moment. No timeline scrubbing. No manual editing. Just the moment.

This is not keyword search. This is structured temporal reasoning.

SceneIQ models video as:

Each scene is represented as a semantic unit:

$$S_i = (T_i,\ E_i,\ O_i,\ M_i,\ R_i,\ I_i)$$

Video Structured Scene Graph.

Video Input
    
Structural Scene Segmentation
    
Motion-Aware Frame Sampling
    
YOLOv8 Object Detection
    
Persistent Multi-Object Tracking
    
Velocity & Motion Modeling
    
Interaction Graph Construction
    
Scene Importance Scoring
    
Semantic Indexing
    
Timestamp-Accurate Retrieval

• HSV histogram comparison
• Structural similarity metrics
• Temporal smoothing
• Narrative-consistent boundary grouping

**Result ** Scene-level units, not raw frames.

Each tracked entity maintains a continuous trajectory:

$$e_j = {(x_t,\ y_t)}_{t=t_1}^{t_2}$$

Track ID continuity enables:

• Long-term identity preservation
• Behavior evolution tracking
• Cross-frame reasoning

Velocity is classified into:

Detects: Goals Celebrations Action spikes High-energy events

Entities become nodes. Spatial proximity and temporal overlap form edges.

Scene graph examples:

person     driving       car
player     holding       trophy
man        speaking_to   woman

Enables semantic reasoning beyond detection.

Scenes are ranked by:

• Motion intensity
• Entity count
• Interaction density

The system surfaces moments that matter.

User query:

"man driving car scene"

Converted into constraints:

• person detected
• vehicle detected
• vehicle_motion > threshold
• spatial overlap: person inside car region

Matched scenes ranked by importance returned with exact timestamp.

Deterministic. Explainable. Local.

Unlike heavy multimodal models, SceneIQ is efficient, transparent, and deployable anywhere.

SceneIQ introduces:

• Structured temporal abstraction
• Persistent identity modeling
• Motion-aware scene importance ranking
• Interaction graph reasoning
• Deterministic semantic retrieval

It bridges classical computer vision and semantic video intelligence.

In the near future, users will not scrub videos. They will query them.

"Last minute winning goal."
"Professor explaining gradient descent."
"Crowd panic moment."

SceneIQ is the engine that makes video searchable by meaning.

SceneIQ doesn't detect frames.
It understands moments.

MIT