Source-linked AI summary
BodyCam-VQA: Enhanced Body-Worn Camera Video Captioning via Multimodal Reasoning and Probe Question Generation
Karish Gupta, Matthew Alex, Alex Li, Yang Wu, Yun-Wei Chu, Kashif Munir, Xiaotian Zhou, Zhengping Ji, Xiaozhong Liu
TL;DR
BWC footage is difficult to process because chaotic, low-quality, rapidly changing scenes and noisy audio can cause VLMs to miss forensic details. BodyCam-VQA uses structured, adaptive visual question answering to enrich captions, with the strongest reported results varying by metric across question-generation pipelines. The framework improves factual accuracy, completeness, and visual enrichment, but current evaluation is limited to 60-second segments and remains vulnerable to hallucinated object attributes.
Problem
BWC videos contain chaotic scenes, low visual quality, rapid interactions, and noisy audio, while VLMs often overlook critical forensic details.
Method
BodyCam-VQA uses adaptive visual question answering and structured reasoning to generate targeted questions that enrich multimodal BWC captions.
Results
BodyCam-VQA improves factual accuracy, completeness, and visual enrichment; Human Factual Accuracy reaches 0.91 for Qwen 3 4B GRPO QA, while Visual Enrichment reaches 0.68 with the DeepSeek-integrated model.
Takeaways & Limitations
Targeted questions increase caption density by prompting models to identify additional people and other context in BWC scenes.
Takeaways & Limitations
The evaluation covers only 60-second segments and does not test long-horizon reasoning or coherent context across linked incident segments.
Abstract
from arXiv · showhide
Police body-worn camera (BWC) footage has emerged as a critical aspect of law enforcement that ensures legal transparency, officer accountability, and the protection of civil rights. However, effectively processing this data remains a significant challenge due to its multimodal video format. BWC videos, in many cases, comprise chaotic scenes with low visual quality, rapid movement/interactions, and high-noise audio that make visual understanding a challenge for even SOTA multimodal models. Current Vision-Language Models (VLMs) frequently overlook critical forensic details, such as the presence of valuable evidence or the latent nuances of suspect-officer interactions, which are vital for fair legal outcomes and civilian/officer safety. To address these limitations, we propose an Adaptive Visual Question Answering (VQA) framework engineered for high-stakes law enforcement. Our framework employs a structured reasoning approach to extract fine-grained visual evidence that traditional captioning systems fail to capture. We experiment with multiple question generation models, including foundation models and fine-tuned open-weight models, to observe performance variation among question generation model implementations. Our results demonstrate that this VQA-driven architecture provides a more reliable, objective, and detailed record of enforcement events, ultimately serving as a powerful tool to protect both law enforcement officers and the public through AI-assisted forensic clarity.
1 Introduction
Body-worn camera footage offers valuable evidence but is difficult to process because chaotic scenes, poor visual quality, rapid interactions, and noisy audio challenge automated understanding. BodyCam-VQA addresses this gap with adaptive visual question answering that extracts overlooked forensic details and improves caption quality.
- Motivation: BWC footage is a large intelligence asset for documenting officer activity and high-stakes encounters, but manual report writing is difficult to scale.Participants who did not review footage were almost four times more likely to make a factual error in reports.
- Motivation: Chaotic visual content, low quality, rapid interactions, and noisy audio make automated understanding of BWC videos challenging.These conditions motivate a VQA system paired with a strong multimodal model for detailed captioning.
- Results: Adding targeted questions changed a caption from describing a single officer to recognizing multi-officer involvement, increasing contextual density.Table 1 presents a qualitative walkthrough of this effect.
- Contributions: The dataset contains 359 annotated 60-second BWC videos with aligned transcripts for multimodal reasoning and analysis.
- Contributions: BodyCam-VQA uses adaptive visual question answering to extract fine-grained forensic evidence that conventional captioning and existing Vision-Language Models often overlook.The framework applies structured reasoning to difficult body-worn camera footage.
2 Related Work
Prior work applies structured multimodal reasoning and visual question answering to improve fine-grained understanding, while public-safety studies show that models struggle with noisy or partially obstructed events.
- Multimodal Reasoning: Structured multimodal reasoning methods use visual instruction tuning, self-distillation, reinforcement learning, and chain-of-thought reasoning to improve VLM alignment and performance.
- Visual Question Answering: Visual question answering elicits targeted, factually grounded information that general-purpose captioning often overlooks.This approach has been used across domains including radiology, pathology, long-form video, and legal consultation.
- VLMs in Public Safety: Public-safety VLM evaluations report stronger event detection when events are clear but weaker detection when events are noisy or partially obstructed.The reviewed systems include optical-flow keyframe selection and anomaly-label scoring.
3.1 Problem Definition
BodyCam-VQA transforms a BWC video and its associated transcript into a VQA-enhanced caption that analyzes the scene, people involved, and critical details.
- 3.1 Problem Definition: Given BWC video V and transcript T, BodyCam-VQA synthesizes a comprehensive caption C capturing scene context, individuals involved, and critical details.
3.2 BodyCam-VQA Mechanism
BodyCam-VQA combines multimodal summarization, structured detail extraction, targeted question generation, and VQA-based caption synthesis. The pipeline uses structured categories and investigative questions to elicit overlooked interactions and objects before producing a refined caption.
- 3.2 BodyCam-VQA Mechanism: The framework comprises multimodal summarization, structured detail extraction, targeted question generation, and VQA-based caption synthesis.
- 3.2 BodyCam-VQA Mechanism: An initial VLM summary provides visual context and a baseline for measuring improvements over transcript-based captioning.
- 3.2.2 Structured Details: Structured details organize observations into scene, entity, action, and semantic categories to ground factual question generation.
- 3.2.3 Question Generation Models: Question-generation quality is treated as central to VQA-enhanced caption quality, with foundation and fine-tuned model architectures evaluated for this role.
- 3.2.6 GRPO Fine-tuned Model: GRPO refines investigative questions using group-relative comparisons and rewards for output format and contextual semantic relevance.
- 3.2.7 VQA and Caption Synthesis: A second VLM pass answers targeted questions, after which the original summary, questions, and answers are synthesized into a caption that can include overlooked interactions and objects.
4 Source Dataset and Video Preprocessing
The study uses COPA body-worn camera incidents and preprocesses selected full-length videos into short, transcript-aligned clips. This segmentation supports manageable VQA processing while preserving synchronized visual and audio information.
- 4 Source Dataset and Video Preprocessing: The dataset samples 40 full-length COPA videos covering diverse police incidents and visual contexts.
- 4 Source Dataset and Video Preprocessing: The preprocessing pipeline segments videos into 60-second clips, producing 359 clips from high-action footage with frames sampled at 1 FPS.
- 4 Source Dataset and Video Preprocessing: Each one-minute segment receives a manually checked, time-aligned transcript generated with WhisperX.
5 Experiment
Experiments show that adding a dedicated VQA stage improves BWC captioning across factual accuracy, completeness, and visual enrichment, while fine-tuned Qwen models remain competitive with closed-source pipelines. Ablations further show distinct trade-offs between reinforcement learning and model scaling, and the case study illustrates improved detail without major hallucinations.
- 5.4 Main Results: Qwen-based models generally exceeded VLM summary baselines and remained competitive with closed-source QA pipelines.Foundation-model pipelines performed better overall, but the performance gap was small.
- 5.4 Main Results: Human Factual Accuracy reached 0.91 with Qwen 3 4B GRPO, Completeness reached 0.65 with Qwen 3 30B SFT, and Visual Enrichment reached 0.68 with DeepSeek QA.These are the strongest reported values for the respective metrics.
- 5.4 Main Results: 0.04-0.13 points better in Completeness and 0.03-0.09 points better in Visual Enrichment were achieved by QA pipelines over baseline models, with denser captions.The comparison covers QA-integrated and non-QA captioning pipelines.
- 5.5 Ablation Study: GRPO increased 4B factual accuracy to 0.91, while SFT achieved higher Completeness than GRPO at 0.57 versus 0.55.The GRPO reward weights for structure and question quality were each set to 0.5.
- 5.5 Ablation Study: Scaling SFT from 4B to 30B increased Completeness to 0.64 but left factual accuracy at 0.86.The reported result separates information density from factual reliability.
- 5.6 Case Study: The fine-tuned caption balanced concision and completeness, capturing multiple officers, flashlight beams, civilian details, and weapon recovery without major hallucinations.The case study contrasts it with a shorter baseline and a more detailed but less accurate QA caption.
6 Conclusion and Future Work
BodyCam-VQA improves caption density, factual accuracy, and visual enrichment, but object and speaker attribution remain difficult in chaotic, first-person footage.
- BodyCam-VQA improves caption density, accuracy, and detail through targeted questions, while residual errors include object misattribution and hallucinated physical attributes.The authors also identify speaker attribution as difficult when video and transcript data are provided separately.
Limitations
The evaluation is limited to 60-second segments and controlled dataset-based scoring, without testing long-horizon reasoning or deployment in operational law-enforcement workflows.
- The pipeline and evaluation use 60-second segments, leaving cross-segment memory and long-horizon reasoning across extended incidents unevaluated.Real-world footage may span minutes to hours and require linking evidence, entities, and events across segment boundaries.
- Operational deployment remains untested, including comparison with officer-written reports, report-drafting integration, and usability studies with investigators or legal reviewers.Such studies are intended to clarify practical time savings, adoption barriers, and remaining accuracy gaps.
Ethical and Societal Implications
The authors frame BodyCam-VQA as supporting legal transparency, accountability, and civil-rights protection while using a publicly released dataset without re-identifying footage or individuals.
- The framework is intended to improve legal transparency, objective accountability, and civil-rights protection through AI-assisted forensic clarity.The authors connect denser, more factually grounded captions with more reliable documentation of high-stakes encounters.
- Experiments use the publicly released COPA dataset and do not collect, process, or re-identify footage or individuals beyond the disclosed source.The authors state that the design is consistent with the ACL Ethics Policy.
A Case Study
The case study compares captions generated from a one-minute body-camera sequence, showing progressively richer scene and interaction detail across model tiers, with the fine-tuned caption performing best.
- The case study’s representative frames mark caption errors in red and notable or key information in bold.
- The QA Caption adds multiple officers, a weapon-recovery statement, movement through the neighborhood, and urgent requests for EMS.Its description expands beyond the initial approach to include the alleyway, street, and officers’ responses to the civilian.
- The captions describe the civilian lying near a garage, with officers illuminating the scene and checking the civilian’s breathing and well-being.The case study includes verbal reassurance, requests for EMS, and descriptions of the civilian’s clothing and surroundings.
- The Fine-Tuned GRPO Caption most richly describes the residential scene, multiple flashlight sources, the civilian’s condition, and officer interactions.It reports the civilian as subdued and verbally responsive while preserving contextual details about vehicles, lighting, and the surrounding property.