AgenticConflictLens: From Long Traffic Videos to Verifiable Conflict Evidence
Organizations: Department of Civil Engineering and Engineering Mechanics Columbia University New York, NY 10027, USA · College of Engineering and Computer Science Florida Atlantic University Boca Raton, 33431, FL, USA
Abstract
Traffic cameras continuously record urban road-user behavior, providing proactive safety evidence beyond crash reports. Yet safety-critical events - conflicts, near misses, unsafe turns, sudden braking, and crosswalk interactions - are rare and brief, buried in hours of routine footage. Existing traffic-video pipelines extract detections, trajectories, and risk scores but process footage uniformly, without adapting to an analyst's query. Vision-language methods improve scene understanding but often lack traffic-specific verification of participant-level conflicts. We introduce AgenticConflictLens, an agentic, query-conditioned framework for efficient, auditable conflict extraction from long fixed-camera traffic videos. Given a natural-language safety query, a Traffic-Aware Controller combines structured traffic memory with specialized conflict-analysis tools to choose where to search, which road users to examine, and which evidence to verify or reject. A coarse-to-fine search retrieves candidate intervals and applies detailed verification selectively. The framework produces conflict records with refined time intervals, participant pairs, semantic regions, trajectories, motion evidence, uncertainty, rejection reasons, and supporting clips. On 96 annotated hours from Bellevue and NYC DOT, AgenticConflictLens achieves the highest conflict-existence F1 and temporal IoU on both datasets: 0.84 F1/0.77 tIoU on Bellevue and 0.80 F1/0.73 tIoU on NYC DOT, while inspecting only 8.7% and 11.9% of video, respectively. These results support scalable, evidence-grounded urban traffic-safety analysis.
Figures & tables
| Bellevue | NYC DOT | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | Exist. F1 | tIoU | Pair Acc. | Inspect % | Avg. Runtime | Exist. F1 | tIoU | Pair Acc. | Inspect % | Avg. Runtime |
| High-risk trajectory intervals Pu and Joshi (2008) | 0.74 | 0.66 | 0.86 | 100.0 | 9.8 | 0.71 | 0.64 | 0.80 | 100.0 | 10.5 |
| Uniform sampled clip conflict search Zhang et al. (2025) | 0.59 | 0.45 | 0.60 | 62.2 | 6.7 | 0.60 | 0.44 | 0.61 | 59.7 | 7.9 |
| Top retrieved conflict clips Wang et al. (2025) | 0.62 | 0.57 | 0.64 | 23.5 | 3.8 | 0.57 | 0.59 | 0.69 | 29.4 | 4.6 |
| Memory-selected conflict clips Yin et al. (2026) | 0.76 | 0.63 | 0.63 | 23.0 | 3.5 | 0.70 | 0.61 | 0.62 | 25.9 | 4.2 |
| Tool-selected conflict clips Lin et al. (2026) | 0.73 | 0.71 | 0.75 | 19.1 | 3.1 | 0.74 | 0.63 | 0.72 | 22.7 | 3.8 |
| Traffic-Analysis Tools | Metrics | ||||||
| QZ | SC | CS | FS | Exist. F1 | tIoU | Pair Acc. | Record Acc. |
| Bellevue | |||||||
| ✓ | ✓ | ✓ | 0.72 | 0.64 | 0.72 | 0.64 | |
| ✓ | ✓ | ✓ | 0.79 | 0.73 | 0.79 | 0.7 | |
| ✓ | ✓ | ✓ | 0.67 | 0.60 | 0.77 | 0.61 | |
| ✓ | ✓ | ✓ | 0.73 | 0.62 | 0.82 | 0.59 | |
| Memory Configuration | Metrics | |||||||
| TS | SP | EB | RR | Context Only | Exist. F1 | tIoU | Pair Acc. | Record Acc. |
| Bellevue | ||||||||
| ✓ | ✓ | ✓ | 0.76 | 0.68 | 0.78 | 0.67 | ||
| ✓ | ✓ | ✓ | 0.77 | 0.67 | 0.80 | 0.66 | ||
| ✓ | ✓ | ✓ | 0.70 | 0.64 | 0.76 | 0.62 | ||
| ✓ | ✓ | ✓ | 0.82 | 0.73 | 0.83 | 0.73 | ||
| Controller Configuration | Metrics | ||||||
| RP | C2F | FB/S | Exist. F1 | tIoU | Record Acc. | Reject % | Inspect % |
| Bellevue | |||||||
| ✓ | ✓ | 0.73 | 0.65 | 0.64 | 73.2 | 19.6 | |
| ✓ | ✓ | 0.76 | 0.66 | 0.66 | 64.4 | 13.9 | |
| ✓ | ✓ | 0.81 | 0.72 | 0.73 | 68.1 | 12.5 | |
| ✓ | ✓ | ✓ | 0.84 | 0.77 | 0.78 | 61.7 | 8.7 |
| Signature | Rule | Threshold |
|---|---|---|
| braking | speed drop approaching closest approach | |
| stopping / stopped | speed below floor, either now or already | |
| accelerating | speed rise approaching closest approach | |
| directional adjustment | heading change around the encounter | |
| evasive motion | either participant adjusts direction | — |
| yielding | one participant slows while the other maintains motion | — |
| Evaluation source | Hours | Conflicts | Queries |
|---|---|---|---|
| Bellevue public evaluation | 60 | 119 | 457 |
| NYC DOT annotated evaluation | 36 | 1008 | 4125 |
| Total annotated evaluation | 96 | 1127 | 4582 |
| Measure | Result |
|---|---|
| Participant-pair agreement | 93.4% |
| Conflict-type agreement | 88.7% |
| Conflict-type Cohen’s | 0.84 |
| Mean temporal IoU | 0.79 |
| Median start-boundary difference | 1.3 s |
| Median end-boundary difference | 1.6 s |
| Family | Number | Share |
|---|---|---|
| Temporally scoped retrieval | 1040 | 22.7% |
| Full-video actor-conditioned search | 1251 | 27.3% |
| Conflict confirmation | 898 | 19.6% |
| Localization and reasoning | 935 | 20.4% |
| Open-ended search | 458 | 10.0% |
| Method | Returned output for comparison | How it is compared |
|---|---|---|
| Top retrieved conflict clips Wang et al. [2025] | Top retrieved conflict clips from hierarchical coarse-to-fine search | tIoU, pair accuracy, efficiency |
| Uniform clip-level conflict search Zhang et al. [2025] | Clips uniformly searched and judged unsafe/conflict by VLM | tIoU, pair accuracy, efficiency |
| Memory-selected conflict clips Yin et al. [2026] | Conflict clips selected from long-video memory entries | tIoU, pair accuracy, efficiency |
| Tool-selected conflict clips Lin et al. [2026] | Conflict clips selected through tool-guided temporal seeking | tIoU, pair accuracy, efficiency |
| High-risk trajectory intervals Pu and Joshi [2008] | Trajectory intervals ranked by surrogate-risk evidence | tIoU, pair accuracy, efficiency |
| AgenticConflictLens | Verified refined conflict intervals and participant-level records | tIoU, pair accuracy, efficiency |
| Retrieval / Localization | Pair Quality | Efficiency | |||||
| Method | Exist. F1 | tIoU | Pair Acc. | Inspect % | Frames | Calls | Runtime |
| Bellevue | |||||||
| High-risk trajectory intervals Pu and Joshi [2008] | 0.74 | 0.66 | 0.86 | 100.0 | 45216 | 0 | 9.8 |
| Uniform sampled clip conflict search Zhang et al. [2025] | 0.59 | 0.45 | 0.60 | 62.2 | 28124 | 224 | 6.7 |
| Top retrieved conflict clips Wang et al. [2025] | 0.62 | 0.57 | 0.64 | 23.5 | 10626 | 96 | 3.8 |
| Memory-selected conflict clips Yin et al. [2026] | 0.76 | 0.63 | 0.63 | 23.0 | 10400 | 88 | 3.5 |
| Method | Scoped | Actor-cond. | Open-ended | Overall |
| Bellevue | ||||
| High-risk trajectory intervals Pu and Joshi [2008] | 0.80 | 0.74 | 0.64 | 0.74 |
| Uniform sampled clip conflict search Zhang et al. [2025] | 0.65 | 0.58 | 0.48 | 0.59 |
| Top retrieved conflict clips Wang et al. [2025] | 0.68 | 0.61 | 0.51 | 0.62 |
| Memory-selected conflict clips Yin et al. [2026] | 0.82 | 0.76 | 0.67 | 0.76 |
| Tool-selected conflict clips Lin et al. [2026] | 0.79 | 0.73 | 0.64 | 0.73 |
| Method | Auto–Ped. | Auto–Cyclist | Auto–Auto | Bus/Truck |
|---|---|---|---|---|
| High-risk trajectory intervals Pu and Joshi [2008] | 0.72 | 0.66 | 0.75 | 0.70 |
| Uniform sampled clip conflict search Zhang et al. [2025] | 0.63 | 0.54 | 0.60 | 0.58 |
| Top retrieved conflict clips Wang et al. [2025] | 0.62 | 0.54 | 0.59 | 0.57 |
| Memory-selected conflict clips Yin et al. [2026] | 0.72 | 0.65 | 0.70 | 0.68 |
| Tool-selected conflict clips Lin et al. [2026] | 0.76 | 0.68 | 0.74 | 0.71 |
| AgenticConflictLens | 0.83 | 0.76 | 0.82 | 0.79 |
Intelligent Transportation Systems
Intelligent Transportation Systems apply sensing, communication, AI, and optimization to make transport networks safer and more efficient.
See all 322 papers