Roadside AI Halts Autonomous Cars at Blind Corners in Scale Trials

Panel 3 - 3: Experimental Evaluation of Infrastructure-Assisted AI for Occluded Driving Scenarios — Ai4

Ai4

Onboard vehicle sensors alone cannot overcome blind urban intersections, requiring static roadside AI perception and wireless telemetry to prevent collisions before entering hazard zones.

AI-assisted summary of Panel 3 - 3: Experimental Evaluation of Infrastructure-Assisted AI for Occluded Driving Scenarios — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 14, 2026, this presentation features Maneesha Vanga, a Graduate Research Assistant and PhD student at the University of Tennessee at Chattanooga. The talk outlines an infrastructure-assisted AI framework designed to help autonomous vehicles safely navigate occluded driving scenarios using roadside perception and lightweight vehicle-to-everything communication.

## TL;DR Maneesha Vanga presents a cooperative driving solution to address the spatial blindness of onboard autonomous vehicle sensors when navigating blind corners and building occlusions. By deploying roadside cameras running YOLO object detection, the infrastructure system identifies pedestrians in blind conflict zones, validates hazards over four consecutive frames, and wirelessly transmits emergency stop commands to oncoming vehicles. Tested across 35 trials on a custom 1/10-scale physical testbed, the system achieved 100% intervention reliability in preventing collisions before safely clearing the vehicle to resume. The project also incorporates 2D and 3D digital twins and a multi-model dashboard to provide a scalable foundation for real-world intelligent transportation networks.

## Atmosphere & style The presentation maintains a crisp, methodical, and focused academic tone centered on pragmatic engineering solutions. Vanga delivers her talk with measured clarity, effectively contrasting everyday human driving instincts with the physical limitations of onboard robotic perception. Watching the session feels practical and innovative, underscoring how accessible scale-model testbeds can rigorously validate critical life-saving autonomous vehicle infrastructure.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) Onboard autonomous vehicle sensors suffer from fundamental line-of-sight limits, leaving them blind to pedestrians and obstacles hidden behind buildings or blind corners. - [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) An infrastructure-assisted AI framework overcomes urban blind spots by using fixed roadside cameras to spot occluded hazards before vehicles approach. - [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) Low-latency wireless communication transmits timely stop commands from roadside AI systems directly to approaching autonomous vehicles. - [01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) While human drivers intuitively decelerate around blind turns, autonomous vehicles relying strictly on onboard perception lack the visibility needed to react safely. - [01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) A complete 1/10-scale physical testbed with mock buildings, roadways, and scaled autonomous vehicles was developed to safely evaluate occluded hazard scenarios. - [01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) Roadside cameras equipped with the YOLO object detection model continuously track designated intersection conflict zones from optimal, unoccluded angles. - [04:05](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=245s) To prevent false triggers, the system enforces a temporal verification check requiring a pedestrian to be detected across four consecutive frames before halting the vehicle. - [04:05](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=245s) Baseline monitor-only trials demonstrated that autonomous vehicles without roadside intervention proceed straight into occluded conflict zones without stopping. - [04:05](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=245s) Active intervention experiments achieved 100% reliability across 20 trials, successfully halting the vehicle before the hazard zone and resuming travel once clear. - [06:27](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=387s) The low-cost 1/10-scale experimental framework was designed to scale directly to full-size municipal infrastructure and real-world connected vehicle deployments. - [06:27](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=387s) Real-time 2D and 3D digital twins pair with multi-model monitoring dashboards to provide fleet operators with comprehensive spatial scene awareness.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) The Challenge of Occlusion and the Infrastructure-Assisted Approach

Maneesha Vanga, a first-year PhD student in computational science at the University of Tennessee at Chattanooga, introduces research conducted alongside her advisors on bolstering the safety of autonomous vehicles (AVs) navigating occluded urban environments. The fundamental dilemma facing autonomous navigation is that onboard vehicle sensors cannot peer around corners or see past solid structures like buildings. When a pedestrian or unexpected hazard enters a crosswalk or intersection behind a building wall, onboard sensors remain completely blind until the vehicle has already turned the corner or entered the hazard area, leaving virtually no reaction time.

To solve this spatial blindness, the research team introduces an infrastructure-assisted artificial intelligence framework. By positioning AI-equipped cameras on static roadside infrastructure overlooking blind corners, the system detects hazards before approaching vehicles reach them. The roadside system processes the visual feed in real time and wirelessly transmits actionable commands directly to the autonomous vehicle, warning it of occluded objects behind buildings so the vehicle can halt prior to entering the conflict zone and only proceed once the hazard has cleared.

```mermaid flowchart TD accTitle: Infrastructure-Assisted AI Intervention Flow subgraph Roadside ["Roadside Infrastructure"] Cam["Roadside Camera Feed"] --> Det["Hazard Detection (YOLO)"] Det --> Ver["Temporal Verification<br/>(4 Consecutive Frames)"] end subgraph Comm ["Wireless Telemetry"] Ver -->|"Stop Command"| CommLink["V2X Communication"] end subgraph Vehicle ["Autonomous Vehicle"] CommLink --> Halt["Timely Vehicle Halt"] Halt --> Clear{"Pedestrian Cleared?"} Clear -->|"Yes"| Resume["Resume Navigation"] Clear -->|"No"| Halt end ```

[01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) Designing a 1/10-Scale Testbed and Roadside Perception

To illustrate the challenge, Vanga draws a comparison to everyday human driving behavior: when approaching a blind intersection, cautious human drivers instinctively "slow down and see what is hidden behind the building." Autonomous vehicles, relying solely on line-of-sight onboard perception, face the exact same physical constraints. Without external visibility, onboard radar, lidar, or cameras cannot compensate for dense physical architecture.

To rigorously investigate this without full-scale hazards, the researchers constructed a complete 1/10-scale physical testbed from scratch. This custom miniature urban environment features simulated infrastructure buildings, roadway signage, marked travel lanes, conflict zones, and scale autonomous vehicles. Fixed roadside cameras were mounted directly to the mock buildings, explicitly angled to eliminate occlusions and monitor blind intersections that vehicle-mounted sensors cannot view. Using the YOLO object detection architecture on the roadside feed, the system continuously analyzes designated conflict zones to spot pedestrians and occluded obstacles long before an oncoming vehicle approaches the turn.

[04:05](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=245s) Experimental Validation: Monitor-Only vs. Active Intervention Trials

The operational logic governing vehicle intervention remains deliberately lightweight and robust. When the roadside perception pipeline detects an occluded object or pedestrian inside a marked conflict zone, the infrastructure AI does not react on a single fleeting false positive. Instead, it tracks the target across four consecutive visual frames. If the pedestrian remains within the conflict zone across those four frames, the system immediately issues a wireless stop command to the approaching autonomous vehicle, keeping it halted until the pedestrian completely clears the hazard area.

The team validated this framework on their physical track through 35 distinct experimental trials:

- **Monitor-Only Baseline Trials:** The infrastructure camera recorded and detected pedestrians within the conflict zone but intentionally withheld stop and clear commands from the approaching vehicle. In these unassisted runs, the autonomous car drove straight toward the blind corner without slowing down, unable to perceive the pedestrian until it was too late. - **Active Intervention Trials:** Across 20 dedicated intervention trials conducted within two designated danger zones, the infrastructure AI identified pedestrians in the conflict zone, validated the hazard, and transmitted an emergency stop command. - **100% Intervention Reliability:** In every single active trial, the vehicle successfully stopped well before entering the conflict zone, remained stationary while the pedestrian was present, and immediately resumed its path once the pedestrian was physically cleared from the track.

[06:27](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=387s) Real-World Translation, Digital Twins, and Future Vision

Although developed on a 1/10-scale autonomous platform, Vanga emphasizes that the underlying architecture, sensor coordination, and low-latency communication model translate directly to full-scale municipal deployment. The physical testbed serves as an accessible, cost-effective proving ground for real-world intelligent transportation systems and connected infrastructure.

To expand the platform's operational capabilities, the team implemented a centralized monitoring dashboard integrating multiple deep learning models. While the live demonstration relies primarily on YOLO for hazard recognition, the dashboard is architected to ingest multi-model perception outputs to inform vehicle maneuvering decisions. Alongside the physical platform, the researchers constructed both two-dimensional and three-dimensional digital twins of the track. These digital representations reflect live physical events in real time, granting human operators and fleet systems an enriched spatial scene understanding. As Vanga concludes, onboard vehicle sensing alone cannot see everything, but layering roadside infrastructure-assisted AI provides the critical redundancy and foresight needed to safeguard complex urban navigation.

## Notable quotes > What happens when an autonomous vehicle cannot see a pedestrian because a building blocks it view? — Maneesha Vanga

> In this talk, I will present an infrastructure-assisted AI framework that helps autonomous vehicles safely navigate occluded driving scenarios using roadside perception and lightweight vehicle communication. — Maneesha Vanga

> [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) The fundamental dilemma facing autonomous navigation is that onboard vehicle sensors cannot peer around corners or see past solid structures like buildings. — Maneesha Vanga

> [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) When a pedestrian or unexpected hazard enters a crosswalk or intersection behind a building wall, onboard sensors remain completely blind until the vehicle has already turned the corner or entered the hazard area, leaving virtually no reaction time. — Maneesha Vanga

> [01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) slow down and see what is hidden behind the building. — Maneesha Vanga

> [04:05](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=245s) In every single active trial, the vehicle successfully stopped well before entering the conflict zone, remained stationary while the pedestrian was present, and immediately resumed its path once the pedestrian was physically cleared from the track. — Maneesha Vanga

> [06:27](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=387s) The physical testbed serves as an accessible, cost-effective proving ground for real-world intelligent transportation systems and connected infrastructure. — Maneesha Vanga

> Attendees will learn how infrastructure-assisted perception can complement onboard sensing, improve safety in challenging environments, and provide a practical foundation for future cooperative autonomous driving systems. — Maneesha Vanga

> If you are interested in AI for transportation, robotics, or intelligent infrastructure, this talk will demonstrate how low-cost experimental platforms can support real-world safety research. — Maneesha Vanga

## Mentioned Maneesha Vanga, The University of Tennessee at Chattanooga, Ai4, YOLO, V2X

## Video description Maneesha Vanga, Graduate Research Assistant, The University Of Tennessee at Chattanooga

What happens when an autonomous vehicle cannot see a pedestrian because a building blocks it view? In this talk, I will present an infrastructure-assisted AI framework that helps autonomous vehicles safely navigate occluded driving scenarios using roadside perception and lightweight vehicle communication. I will explain how we developed and experimentally evaluated this approach on a 1/10-scale autonomous vehicle platform, demonstrating reliable pedestrian detection and timely vehicle intervention. Attendees will learn how infrastructure-assisted perception can complement onboard sensing, improve safety in challenging environments, and provide a practical foundation for future cooperative autonomous driving systems. If you are interested in AI for transportation, robotics, or intelligent infrastructure, this talk will demonstrate how low-cost experimental platforms can support real-world safety research.

[Source](https://www.youtube.com/watch?v=MV7V_rk7XQo)

中文简报

缩比试验显示路侧AI可在盲区弯道叫停自动驾驶汽车 仅凭车载传感器无法克服城市交叉路口的视野盲区，需要静态路侧AI感知与无线遥测技术在车辆进入危险区域前防止碰撞。 Maneesha Vanga 提出了一种协同驾驶解决方案，以解决自动驾驶车辆车载传感器在盲区弯道和建筑物遮挡下的空间盲区问题。该基础设施系统通过部署运行 YOLO 目标检测算法的路侧摄像头，识别盲区冲突区域内的行人，在连续四帧确认危险后，通过无线方式向驶来的车辆传输紧急制动指令。在定制的 1/10 缩比物理试验台上进行的 35 次试验中，该系统在防止碰撞方面的干预可靠性达到 100%，并在确认安全后允许车辆恢复行驶。该项目还集成了 2D 与 3D 数字孪生以及多模型仪表板，为现实世界智能交通网络的构建提供了可扩展的基础。 由于车载传感器受限于视线范围无法看透建筑物拐角，研究人员提出了路侧基础设施协同 AI 架构，利用路侧固定视角提前捕捉遮挡盲区的行人和障碍物。 (0:04) 研究团队打造了包含微缩建筑与道路的 1/10 比例物理测试平台，并在路侧部署运行 YOLO 目标检测模型的摄像头，对路口冲突区域进行无死角监控。 (1:53) 系统采用连续 4 帧检测确认的时序验证机制以防止误触发，在 20 次主动干预测试中实现了 100% 的成功刹车率，并在行人离开后自动恢复行驶。 (4:05) 该架构支持直接迁移至全尺寸真实交通场景，团队还构建了配套的 2D 与 3D 数字孪生系统，通过路侧协同感知为城市自动驾驶提供冗余安全保障。 (6:27)

Reading summary

Researchers at the University of Tennessee at Chattanooga propose an infrastructure-assisted artificial intelligence framework to overcome line-of-sight limitations that leave autonomous vehicle onboard sensors blind at occluded urban intersections. Onboard Perception Limits at Blind Intersections (0:04) Autonomous vehicle navigation remains constrained by physical line-of-sight barriers such as buildings and roadside architecture. Because onboard radar, lidar, and cameras cannot peer around corners, an autonomous vehicle approaching a blind intersection cannot detect pedestrians or unexpected obstacles in its path until it has already turned, leaving virtually no reaction time to avert a collision. Roadside AI Perception and Temporal Verification (4:05) To provide early hazard visibility, the proposed architecture positions static, AI-equipped roadside cameras overlooking blind zones. The system monitors designated conflict areas in real time using the YOLO object detection framework. To prevent erratic maneuvers from fleeting false positives, the infrastructure logic tracks targets across four consecutive visual frames; only when an obstacle remains in the zone across all four frames does the roadside unit wirelessly transmit an emergency stop command to approaching vehicles, keeping them halted until the hazard clears. Empirical Validation on a 1/10-Scale Testbed (1:53) The researchers evaluated their system using a custom 1/10-scale physical urban testbed equipped with mock buildings, travel lanes, conflict zones, and scale vehicles. Across 35 experimental runs, the team compared unassisted and assisted driving responses. In baseline monitor-only runs where infrastructure warnings were intentionally suppressed, vehicles drove directly toward blind corners at cruising speed, unable to detect pedestrians in time. In 20 active intervention trials across two danger zones, the infrastructure AI achieved 100% intervention reliability, successfully halting the scale vehicle prior to the conflict zone every time and resuming travel once the pedestrian cleared the track. Digital Twins and Municipal Deployment Scalability (6:27) While demonstrated on a miniature track, presenter Maneesha Vanga asserts that the underlying sensor coordination and low-latency wireless communication model are directly transferable to full-scale intelligent transportation systems. To support fleet oversight, the platform incorporates real-time two-dimensional and three-dimensional digital twins alongside a centralized dashboard designed to ingest multi-model perception data beyond YOLO to guide vehicle maneuvering. The study demonstrates that vehicle-mounted sensors alone cannot resolve severe urban occlusions, framing infrastructure-based perception as a necessary safety redundancy. The researchers argue that the 1/10-scale testbed offers a cost-effective proving ground for municipal deployment, though real-world effectiveness will depend on reliable low-latency communication and physical infrastructure investment at blind corners.

精炼阅读摘要

田纳西大学查塔努加分校研究团队提出一种基于路侧基础设施辅助的人工智能框架，通过在盲区路口部署静态路侧感知与车联网无线控制，解决自动驾驶汽车车载传感器受建筑物遮挡而无法提前预警的行车安全难题。 视距受限盲区与路侧协同构想 (0:04) 在复杂的城市驾驶环境中，自动驾驶汽车单凭车载雷达、激光雷达或摄像头等传感设备，无法穿透建筑物或绕过转角捕捉盲区内的危险。研究员Maneesha Vanga指出，普通人类驾驶员在驶向视野盲区路口时通常会本能地减速观察，而完全依赖车载视距感知的自动驾驶车辆受制于物理结构遮挡，如果在行人已处于拐角后方冲突区时才完成识别，车辆将几乎没有制动反应时间。 针对该空间盲区难题，该团队设计了基础设施辅助人工智能架构。通过在俯瞰盲区交叉口的路侧静态建筑上安装配备AI算法的摄像头，系统能够在来车接近之前提前感知隐蔽目标，并将视觉流数据进行实时计算，以低延迟无线通信方式直接向自动驾驶车辆下达行动指令，指示车辆在进入危险区前安全停下，待行人离开后再行通过。 缩比物理测试平台与视觉感知设计 (1:53) 为了在避免全尺寸实车风险的前提下进行严谨验证，该团队从零开始搭建了一套完整的十分之一比例（1/10-scale）物理微缩测试环境。该测试平台包含模拟城市建筑物、道路标线、交通指示牌、指定冲突区域以及缩比自动驾驶车辆，直观复刻城市死角路况。 研发人员将固定路侧摄像头直接架设于微缩建筑表面，并特别调整安装角度，确保完全消除视野遮挡，从而监控车载设备视距之外的死角路口。在路侧感知管道中，系统部署了YOLO目标检测模型，对划定的道路冲突区域实施连续监测，在来车尚未驶近转弯处之前，即可精确识别出位于隐蔽区域的行人与障碍物。 连续帧校验与主动干预实验验证 (4:05) 为保障车辆干预决策的稳定性并滤除偶发性误报，该基础设施AI系统引入了轻量化的时间连续性校验机制。系统在道路冲突区域检测到行人后并不会依据单一帧图像立即反应，而是对其进行跨越连续4个视频帧的持续追踪；唯有目标在连续4帧内均处于冲突区，系统才会立即向接近的自动驾驶车辆发出无线急停指令，并维持制动直至行人完全移出该危险区域。 研究团队在该测试跑道上累计执行了35次对照实验。在仅监测的基准测试中，路侧系统检测到冲突区行人后刻意不向来车发送制动指令，测试车辆因自身无法透视建筑物而全速驶向盲区；而在包含两个指定危险区域的20次主动干预测试中，路侧AI均成功捕获目标并发出停止信号，车辆在进入冲突区前实现100%可靠停车，待行人离开后立即自主恢复行驶。 数字孪生映射与现实城市落地展望 (6:27) 尽管该验证方案目前运行于十分之一比例的试验台，但Vanga强调，其背后的系统架构、多传感器协同逻辑以及低延迟通信机制可直接外推并迁移至全尺寸的市政道路基础设施中。该物理试验场为智能交通系统（ITS）与车路协同网络提供了一个高性价比且易于迭代的原型验证环境。 为了强化全局监控与决策能力，团队还构建了集中式监控仪表板，虽然现场演示主要依靠YOLO算法识别危险，但该后台在架构上支持吸纳多种深度学习模型的感知输出，用以辅助复杂的行车机动决策。此外，研究人员针对测试跑道同步构建了二维与三维数字孪生系统，实时镜像物理环境中的动态事件，为人类运营人员和车队调度系统提供更丰富的全局空间态势感知。 研究表明，仅依靠车载传感器无法在视距受阻的复杂城市街道中实现全景感知，引入路侧基础设施辅助AI能够提供关键的安全冗余与前瞻预警；不过，该研究目前仍建立在受控缩比测试平台之上，向全尺寸真实城市道路规模化部署过程中的通信时延、复杂天气适应性及多车协同调度仍有待进一步工程验证。

Detailed notes — zh

## Background 本内容由 Ai4 于 2026 年 8 月 14 日在 YouTube 平台发布。视频主要呈现了来自田纳西大学查塔努加分校的研究生 Maneesha Vanga 的学术演讲，探讨如何利用路侧基础设施 AI 与轻量级车路通信，解决自动驾驶车辆在建筑物盲区下的视线遮挡感知难题，并在 1/10 缩比实验平台上进行了系统验证。 ## TL;DR 研究针对自动驾驶车辆面临建筑物视线遮挡、车载传感器存在感知盲区的痛点，提出了基于路侧基础设施 AI 的车路协同辅助感知方案。团队构建了包含建筑物盲区的 1/10 物理缩比实验平台，利用部署在路侧高处的摄像头配合 YOLO 目标检测模型实时监测冲突区域。系统引入连续 4 帧目标确认防抖机制，并在 35 轮实车测试（含 20 次主动干预测试）中达成了 100% 的成功识别与安全刹停，同时结合 2D/3D 数字孪生系统展现了车路协同架构向现实智慧城市落地的广阔前景。 ## Atmosphere & style 整场演讲严谨干练且极具学术工科气息，主讲人条理清晰、语速平稳地拆解了从痛点分析、算法设计到硬件实测的全流程。伴随着缩比物理赛道演示与决策流程图的直观展示，观众能清晰感受到从理论设想到工程落地的踏实感，展现出年轻学者在自动驾驶安全前沿探索中的专注与自信。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) 自动驾驶车辆在城市环境中极易受到建筑物遮挡，单车车载传感器受视线几何限制存在严重的感知盲区与碰撞风险。 - [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) 提出基于基础设施辅助 AI（车路协同）的技术方案，通过路侧盲区感知与轻量化通信通道，向车载系统下发主动干预指令。 - [01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) 构建 1/10 比例的物理实验沙盘与自主小车平台，在安全、低成本的环境下高保真复现遮挡交叉路口场景。 - [01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) 在建筑物高处架设路侧摄像头建立俯视视角，彻底打破单车感知的视距死角，全面覆盖潜在冲突区域。 - [01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) 路侧端运行 YOLO 目标检测模型，能够在车载系统未发现目标前毫秒级识别并定位盲区横穿的行人。 - [04:05](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=245s) 引入连续 4 帧稳定检测确认的防抖决策逻辑，有效过滤瞬时误检，保障主动制动指令下发的可靠性与平顺性。 - [04:05](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=245s) 物理实验对比显示，仅监控基线模式下车辆会盲目冲向冲突区，而主动干预模式在 20 次盲区行人测试中达成 100% 成功刹停与安全放行。 - [06:27](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=387s) 开发 2D 与 3D 数字孪生系统与监控看板，实时映射物理实体轨迹，并预留多深度学习模型接入接口以丰富环境语义。 - [06:27](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=387s) 缩比平台验证的跨视角协同与低延迟通信底层逻辑，可平滑推广至真实智慧城市与网联自动驾驶（V2X）场景中。 - [06:27](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=387s) 单车硬件感知无法完全消除盲区死角，将路侧 AI 作为“第二双眼睛”是实现高安全级别网联自动驾驶的重要演进方向。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) 遮挡场景下的自动驾驶困境与车路协同设想 来自田纳西大学查塔努加分校（University of Tennessee at Chattanooga）计算科学专业的博士一年级研究生 Maneesha Vanga 带来了关于车路协同 AI 的研究分享。她指出，自动驾驶车辆（AV）在城市街区行驶时，常常面临视线被建筑物遮挡的严峻挑战。当车辆接近视野受限的盲区交叉路口时，仅依靠车载传感器很难察觉隐藏在建筑物背后的危险。针对这一痛点，该研究探索了利用路侧基础设施 AI（Infrastructure-Assisted AI）来辅助车载系统，及时识别遮挡区域内的潜在危险，以此大幅提升自动驾驶的安全性。 整个研究的技术路线围绕“感知—通信—决策”展开：部署在路侧基础设施上的 AI 摄像头负责监视盲区，一旦检测到建筑物后方存在行人或障碍物等危险目标，便会通过轻量化通信通道向车载自动驾驶系统发送警报及控制指令。车载系统接收到“建筑物后存在遮挡物体”的信息后，会主动制动停车，直到目标彻底离开冲突区域（Conflict Zone），车辆再安全恢复通行。 ```mermaid flowchart TD accTitle: 车路协同路侧感知与主动干预决策流程图 subgraph 路侧基础设施感知 A["路侧固定摄像头捕获视野"] --> B["YOLO 目标检测模型识别盲区行人"] B --> C{"连续 4 帧检测确认冲突区存在目标?"} end subgraph 车路协同通信与决策 C -- "是" --> D["路侧 AI 发送 Stop 指令"] D --> E["车载 AV 提前减速并驻车等待"] C -- "否 / 目标离开" --> F["路侧 AI 发送 Clear 指令恢复通行"] F --> G["车载 AV 继续按既定路线行驶"] end ``` --- ### [01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) 1/10 缩比实验平台搭建与路侧 YOLO 目标感知 为了在安全、低成本的环境下复现并验证这一设想，研究团队从零构建了一套 1/10 比例的微缩真实驾驶测试沙盘。沙盘内精确布置了微缩建筑物、道路交通标志以及具备自主行驶能力的 1/10 比例自动驾驶小车平台。人类驾驶员在驶向视线受阻的路口时，通常会习惯性减速并探头观察建筑物后的死角，但当前的自动驾驶车载传感器受限于安装位置与视线距离（Line of Sight），在遇到实体建筑遮挡时会产生严重的感知盲区，无法提前发现拐角处的行人。 针对这些盲区盲道，研究团队将多台路侧摄像头架设在微缩建筑物的高处，专门覆盖赛道中的遮挡冲突区域。在感知算法层面，系统部署了 YOLO 目标检测模型，用于实时分析路侧摄像头捕获的画面。当路侧画面中出现横穿的行人模型时，YOLO 能够在车载传感器完全看不到行人的情况下，率先识别出行人目标并定位其冲突风险。 - 物理测试环境包含按比例缩放的建筑物盲区、标准交通标志及 1/10 比例全自主行驶平台。 - 路侧摄像头专为覆盖车载视线死角而布置，从俯视视角打破单车感知的几何局限。 - 感知后端运行 YOLO 目标检测算法，实现毫秒级的行人识别与潜在冲突判定。 --- ### [04:05](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=245s) 干预决策逻辑与 35 轮实车对比测试 系统的控制决策逻辑设计得清晰而直接。当自动驾驶小车向冲突区驶来时，路侧 AI 会持续监控该危险区域；为防止单帧误检引发不必要的急刹，路侧端设置了防抖逻辑——只有在连续 4 帧图像中均稳定检测到冲突区内存在行人时，AI 才会向车载系统发出制动指令（Stop Command）。车载小车在驶入盲区路口前即执行减速停车，静止等待，直到行人离开危险区域，系统确认道路出清后才允许车辆重新启程。 为了验证路侧主动干预的有效性，团队在物理赛道上累计进行了 35 轮实测对比，分为“仅监控”（Monitor-only）基线模式与“主动干预”（Active Intervention）模式： - **仅监控基线测试**：摄像头正常检测冲突区域中的行人，但路侧系统不向小车下发任何停止或放行指令。小车在盲区前无法提前感知行人，在多轮测试中直接保持原速冲向冲突区，暴露了极大的碰撞风险。 - **主动干预测试**：团队在两个危险冲突区域共执行了 20 次行人侵入测试。在全部 20 次测试中，路侧 AI 均实现 100% 成功识别，并准时下发停车指令让小车在路口盲区前平稳刹停；当实验人员将行人模型移出冲突区后，小车均能顺畅恢复巡航。 --- ### [06:27](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=387s) 数字孪生集成与车路协同技术的现实落地 Maneesha Vanga 强调，尽管目前的验证是在 1/10 缩比平台上完成的，但其背后的车路协同底层逻辑、低延迟通信机制与跨视角感知算法完全可以无缝推广到真实世界的智慧城市与车联网（V2X）场景中。 此外，团队还围绕实验平台开发了一套综合监控看板与数字孪生系统。虽然现阶段目标检测主要基于 YOLO 模型，但系统架构已预留多模型接口，未来可集成更多深度学习模型以丰富环境语义信息。研究人员还构建了二维与三维数字孪生（2D and 3D Digital Twin），将物理赛道上的车辆位置、行人动态和路侧视角实时映射到数字化屏幕上，极大增强了全局场景理解能力。Vanga 在总结中指出，"自动驾驶车辆不可能看清所有角落"，但通过引入基础设施辅助 AI 作为第二双眼睛，能够显著增强单车的态势感知与防撞预警能力，为未来高安全性的网联自动驾驶奠定坚实基础。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) 自动驾驶车辆在城市街区行驶时，常常面临视线被建筑物遮挡的严峻挑战。当车辆接近视野受限的盲区交叉路口时，仅依靠车载传感器很难察觉隐藏在建筑物背后的危险。 — Maneesha Vanga > [00:04](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=4s) 整个研究的技术路线围绕“感知—通信—决策”展开：部署在路侧基础设施上的 AI 摄像头负责监视盲区，一旦检测到建筑物后方存在行人或障碍物等危险目标，便会通过轻量化通信通道向车载自动驾驶系统发送警报及控制指令。 — Maneesha Vanga > [01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) 人类驾驶员在驶向视线受阻的路口时，通常会习惯性减速并探头观察建筑物后的死角，但当前的自动驾驶车载传感器受限于安装位置与视线距离，在遇到实体建筑遮挡时会产生严重的感知盲区。 — Maneesha Vanga > [01:53](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=113s) 当路侧画面中出现横穿的行人模型时，YOLO 能够在车载传感器完全看不到行人的情况下，率先识别出行人目标并定位其冲突风险。 — Maneesha Vanga > [04:05](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=245s) 为防止单帧误检引发不必要的急刹，路侧端设置了防抖逻辑——只有在连续 4 帧图像中均稳定检测到冲突区内存在行人时，AI 才会向车载系统发出制动指令。 — Maneesha Vanga > [04:05](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=245s) 在全部 20 次测试中，路侧 AI 均实现 100% 成功识别，并准时下发停车指令让小车在路口盲区前平稳刹停；当实验人员将行人模型移出冲突区后，小车均能顺畅恢复巡航。 — Maneesha Vanga > [06:27](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=387s) 尽管目前的验证是在 1/10 缩比平台上完成的，但其背后的车路协同底层逻辑、低延迟通信机制与跨视角感知算法完全可以无缝推广到真实世界的智慧城市与车联网场景中。 — Maneesha Vanga > [06:27](https://www.youtube.com/watch?v=MV7V_rk7XQo&t=387s) 自动驾驶车辆不可能看清所有角落，但通过引入基础设施辅助 AI 作为第二双眼睛，能够显著增强单车的态势感知与防撞预警能力，为未来高安全性的网联自动驾驶奠定坚实基础。 — Maneesha Vanga ## Mentioned Maneesha Vanga, The University Of Tennessee at Chattanooga, Ai4, YOLO, V2X, 1/10 缩比自动驾驶平台, 数字孪生系统 ## Video description Maneesha Vanga, Graduate Research Assistant, The University Of Tennessee at Chattanooga What happens when an autonomous vehicle cannot see a pedestrian because a building blocks it view? In this talk, I will present an infrastructure-assisted AI framework that helps autonomous vehicles safely navigate occluded driving scenarios using roadside perception and lightweight vehicle communication. I will explain how we developed and experimentally evaluated this approach on a 1/10-scale autonomous vehicle platform, demonstrating reliable pedestrian detection and timely vehicle intervention. Attendees will learn how infrastructure-assisted perception can complement onboard sensing, improve safety in challenging environments, and provide a practical foundation for future cooperative autonomous driving systems. If you are interested in AI for transportation, robotics, or intelligent infrastructure, this talk will demonstrate how low-cost experimental platforms can support real-world safety research. [Source](https://www.youtube.com/watch?v=MV7V_rk7XQo)

Source quotes

[0:04] Maneesha Vanga: 自动驾驶车辆在城市街区行驶时，常常面临视线被建筑物遮挡的严峻挑战。当车辆接近视野受限的盲区交叉路口时，仅依靠车载传感器很难察觉隐藏在建筑物背后的危险。 [0:04] Maneesha Vanga: 整个研究的技术路线围绕“感知—通信—决策”展开：部署在路侧基础设施上的 AI 摄像头负责监视盲区，一旦检测到建筑物后方存在行人或障碍物等危险目标，便会通过轻量化通信通道向车载自动驾驶系统发送警报及控制指令。 [1:53] Maneesha Vanga: 人类驾驶员在驶向视线受阻的路口时，通常会习惯性减速并探头观察建筑物后的死角，但当前的自动驾驶车载传感器受限于安装位置与视线距离，在遇到实体建筑遮挡时会产生严重的感知盲区。 [1:53] Maneesha Vanga: 当路侧画面中出现横穿的行人模型时，YOLO 能够在车载传感器完全看不到行人的情况下，率先识别出行人目标并定位其冲突风险。 [4:05] Maneesha Vanga: 为防止单帧误检引发不必要的急刹，路侧端设置了防抖逻辑——只有在连续 4 帧图像中均稳定检测到冲突区内存在行人时，AI 才会向车载系统发出制动指令。 [4:05] Maneesha Vanga: 在全部 20 次测试中，路侧 AI 均实现 100% 成功识别，并准时下发停车指令让小车在路口盲区前平稳刹停；当实验人员将行人模型移出冲突区后，小车均能顺畅恢复巡航。 [6:27] Maneesha Vanga: 尽管目前的验证是在 1/10 缩比平台上完成的，但其背后的车路协同底层逻辑、低延迟通信机制与跨视角感知算法完全可以无缝推广到真实世界的智慧城市与车联网场景中。 [6:27] Maneesha Vanga: 自动驾驶车辆不可能看清所有角落，但通过引入基础设施辅助 AI 作为第二双眼睛，能够显著增强单车的态势感知与防撞预警能力，为未来高安全性的网联自动驾驶奠定坚实基础。

Source quotes

Maneesha Vanga: What happens when an autonomous vehicle cannot see a pedestrian because a building blocks it view? Maneesha Vanga: In this talk, I will present an infrastructure-assisted AI framework that helps autonomous vehicles safely navigate occluded driving scenarios using roadside perception and lightweight vehicle communication. [0:04] Maneesha Vanga: The fundamental dilemma facing autonomous navigation is that onboard vehicle sensors cannot peer around corners or see past solid structures like buildings. [0:04] Maneesha Vanga: When a pedestrian or unexpected hazard enters a crosswalk or intersection behind a building wall, onboard sensors remain completely blind until the vehicle has already turned the corner or entered the hazard area, leaving virtually no reaction time. [1:53] Maneesha Vanga: slow down and see what is hidden behind the building. [4:05] Maneesha Vanga: In every single active trial, the vehicle successfully stopped well before entering the conflict zone, remained stationary while the pedestrian was present, and immediately resumed its path once the pedestrian was physically cleared from the track. [6:27] Maneesha Vanga: The physical testbed serves as an accessible, cost-effective proving ground for real-world intelligent transportation systems and connected infrastructure. Maneesha Vanga: Attendees will learn how infrastructure-assisted perception can complement onboard sensing, improve safety in challenging environments, and provide a practical foundation for future cooperative autonomous driving systems. Maneesha Vanga: If you are interested in AI for transportation, robotics, or intelligent infrastructure, this talk will demonstrate how low-cost experimental platforms can support real-world safety research.

Original transcript

[0:04] Hello everyone. Uh my name is Manisha Wanga. I'm  a first year PhD student in computational science. [0:11] I'm from University of Tennessee Chhattanooga  and those are my advisers. So today I would [0:18] like to share how infrastructure assisted AI can  help or improve the safety of autonomous vehicles [0:31] driving in oluded scenarios. So we specifically  investigated how an infrastructure AI [0:41] can help the autonomous vehicle to detect  the hazards in occluded urban scenarios. [0:51] I will begin with the road map the problem. So as  we know that autonomous vehicles they struggle to [1:02] see beyond the buildings and the solution we  came up with using an infrastructure assisted AI [1:13] camera that detects the hazards and using  that information we will send the commands [1:20] to the onboarded AV saying there's a occluded  object behind this building. So based on that [1:27] information, the AV should stop and then clear  when the object is cleared from the config zone. [1:38] And I'll be introducing I will  also be presenting the evolution [1:44] the results and the impact of  this research in a real world. [1:53] So firstly let's imagine if there if we  are driving in a occluded uh intersection. [2:02] So even as a human drivers we slow down  and see what is hidden behind the building. [2:09] So the same challenge happens to the autonomous  vehicles. They struggle to see behind the [2:18] buildings. So we came up with solution where  we built a scaledown environment from scratch [2:27] with all the infrastructure buildings, road signs  and scaled down autonomous vehicle platforms. [2:35] So with that we mounted few cameras  to our infrastructure buildings [2:42] specifically to cover those occluded scenarios  on the track. As you can see the cameras [2:51] areas are covered on the  particular occluded scenarios. [3:00] So the problem as I as I told Even though  autonomous vehicles they have their [3:09] onboard sensors but the visibility is very less  with only using its onboard sensors and also [3:19] the buildings can create uh blind spots  and pedestrians cannot be detected [3:27] until unless it sees with its onboard sensors.  So here you can see in the camera one is the [3:36] view from the infrastructure camera using a model  yolo. It is able to detect there is a person [3:46] pedestrian in that particular  view. So using that information [3:53] we'll send that information to the onboarded  AV even before it sees it with onboard sensors. [4:05] So the experimental platform is like the  framework is so simple. We detect the [4:14] occluded objects and pedestrians in a particular  conflict zone. And when the vehicle is [4:21] approaching the that uh conflict zone, the a the  infrastructure AI, it sees until four consecutive [4:33] frames. Even though pedestrian is still  there, the AI sends a stop command to the [4:42] onboarded AV. So then the AV should stop until  the pedestrian is cleared from the conflict zone. [4:53] So this is the physical the test we have  done on our track. We did like 35 trails [5:03] with monitor only. So even though the camera sees  a pedestrian in the conflict zone, it do not send [5:12] any stop command or clear command to the onboarded  AV because it's just a monitor only view. [5:20] You can see the car approaching the  conflict zone, but it still didn't stop. [5:29] In this video, you can see this is a view where  the infrastructure AI sends a command to the [5:38] onboarded vehicle. So, the vehicle is stopping  even before it is coming to the config zone. [5:49] So you can see the vehicle stopped [5:52] and when I removed the pedestrians from that  conflict zone the vehicle resumed its path. [6:01] As I told we did 35 trails 20 active  intervention where in two danger zones 20 times [6:10] the potention were present in that conflict  zone. every trail the infrastructure AI is [6:19] able to successfully detected the pedestrian and  we were able to send the stop command to the AV. [6:27] So what we can take from this research to our  real world. We did this although we did this on [6:35] a 1x10 scale AV platform but the core concept and  behind this is totally applicable to a real world. [6:48] With that we also implemented a dashboard  with few AI deep learning models. [6:55] For now, we are only using YOLO to detect  the objects. But in future, we'll be able to [7:02] use all those models to gather information from  a track and send that based on that information, [7:11] the onboarded AV should take a decisions and also  we implemented two-dimensional digital twin and [7:20] three-dimensional digital twin. So  to give more scene understanding of [7:26] whatever is happening on a physical  track to a digital screen. So [7:35] autonomous vehicles they cannot see everything  but adding a layer with infrastructure assisted AI [7:45] can improve safety and alertness  in onboarded AS. Yeah. Thank you.