Eyes on the Road: State-of-the-Art Video Question Answering Models Assessment for Traffic Monitoring Tasks

TL;DR

This work addresses the challenge of deploying VideoQA for traffic monitoring by benchmarking SOTA multimodal models on a mix of real-world Tempe AZ footage and CARLA synthetic sequences. The authors deploy a comprehensive evaluation framework that treats GPT-4o as an objective semantic evaluator and employs compositional consistency metrics to assess reasoning across detection, temporal reasoning, and decomposition queries. VideoLLaMA-2 emerges as the strongest model (average $57\%$ accuracy) particularly in compositional and long-horizon reasoning, but all models struggle with precise multi-object tracking and temporal coherence, underscoring persistent architectural gaps. The study demonstrates the promise of VideoQA for incident detection and traffic management while highlighting targeted avenues for improvement and providing open-source tools to advance future research and deployment in intelligent transportation systems.

Abstract

Recent advances in video question answering (VideoQA) offer promising applications, especially in traffic monitoring, where efficient video interpretation is critical. Within ITS, answering complex, real-time queries like "How many red cars passed in the last 10 minutes?" or "Was there an incident between 3:00 PM and 3:05 PM?" enhances situational awareness and decision-making. Despite progress in vision-language models, VideoQA remains challenging, especially in dynamic environments involving multiple objects and intricate spatiotemporal relationships. This study evaluates state-of-the-art VideoQA models using non-benchmark synthetic and real-world traffic sequences. The framework leverages GPT-4o to assess accuracy, relevance, and consistency across basic detection, temporal reasoning, and decomposition queries. VideoLLaMA-2 excelled with 57% accuracy, particularly in compositional reasoning and consistent answers. However, all models, including VideoLLaMA-2, faced limitations in multi-object tracking, temporal coherence, and complex scene interpretation, highlighting gaps in current architectures. These findings underscore VideoQA's potential in traffic monitoring but also emphasize the need for improvements in multi-object tracking, temporal reasoning, and compositional capabilities. Enhancing these areas could make VideoQA indispensable for incident detection, traffic flow management, and responsive urban planning. The study's code and framework are open-sourced for further exploration: https://github.com/joe-rabbit/VideoQA_Pilot_Study

Figures (10)

• Figure 1: An example scenario of a question posed by a traffic analyst and answered by a VideoQA model is based on a traffic video sequence captured at a real-world four-way intersection.
• Figure 2: Traffic Video Data Sources - (a) Real-world traffic video captured at a four-way intersection in Tempe, AZ. (b) Sample video frame from the Tempe intersection, showing real-world traffic interactions. (c) Simulated video frame from CARLA, depicting a four-way intersection with varied vehicle and pedestrian activity in a controlled environment.
• Figure 3: Sample Scenarios from traffic Videos - Sequence 1 (top) represents a real-world scenario captured at a four-way intersection on S. Mill Avenue in Tempe, AZ, featuring a cyclist. Sequence 2 (middle) depicts cars turning onto a bike lane, captured in another real-world setting. Sequence 3 (bottom) is synthesized using CARLA, illustrating a simulated four-way intersection with diverse vehicle and pedestrian activity.
• Figure 4: Shows the architectures of three open-source models: (a) InternVL chen2023internvlchen2024far, (b) VideoLLaMA-2 damonlpsg2023videollama, and (c) the general architecture of the LLAVA-One family li2024llava.
• Figure 5: (a) Shows an extensive summary of a $45mins$ lecture provided by GPT-$4o$, it includes subjection for each topic discussed in the video gpt4o2024, (b) represents the Gemini Pro VideoQA capabilities the movie Sherlock Jr. of $44mins$ is passed. The superiority of the model to answer questions as well as recognize hand-drawn images and justify the answer by providing the exact frame as validation figure referenced from geminiteam2024geminifamilyhighlycapable.