编者按：电子工程专辑是全球最大的电子工程独立资源平台，面向设计工程师、设计管理者、技术专家及行业高管，提供电子行业相关的技术动态、深度分析与行业观点，涵盖影响产业发展的技术与商业议题。欢迎访问原文。为方便阅读，我们为您全文转载中英文内容。
Notes:EETimes is the largest independent global engineering resource for design engineers, design managers, technologists and executives, offering news, analysis and perspectives on the technology and business issues that shape the electronics industry. And you are welcome to pay visit to the full article via below links:
1.https://www.eetimes.com/reshaping-the-eye-of-ai-how-will-hvs-transform-ai-machine-vision-in-2026/
2.https://www.eet-china.com/news/202602069667.html
作者：黄烨锋，资深产业分析师，电子工程专辑
前两年，电子工程专辑就一直在关注“事件视觉传感器（event-based vision sensor）”：相比传统的CIS（CMOS图像传感器），它只捕捉画面中动起来的部分，基于事件触发，记录连续事件、等效上万帧每秒的帧率是其特色。这类传感器在汽车、医疗、工业制造，乃至消费电子等领域均有应用潜力。
最近在与锐思智芯的对话中，锐思智芯创始人兼CEO邓坚谈到事件感知对当代科技的系统性价值：“AI时代需要的是多维度、多模态数据——比如自动驾驶，不仅需要x和y轴的两维平面图像数据；z轴的距离信息；还需要颜色、光谱信息；更重要的是时间和运动维度...这些构成了当代视觉系统的几大要素。”
有了多维度的数据，机器才能准确感知世界，并做出合适的决策。“就像平面图像有传统的图像传感器，距离有ToF或其他测距传感器，光谱传感器能够提供颜色信息。在时间、运动多个维度，信息当然可以通过传统视频拍摄的方式来获取——但视频的问题在于帧率不高、无效和重复数据多，很难有效提取有价值信息。”
而事件感知技术，通过仿人眼的工作模式，对电脉冲进行串联、对变化的信息做感知，与传统图像传感器的静态信息做互补。因此，锐思智芯从2019年创立以来就在做一种融合视觉传感器（HVS，hybrid vision sensor），将事件感知与图像感知融合，“将静态与变化两种感知能力，融合在一个像素之中。未来搭配距离传感器、光谱传感器等，将视觉所需的要素都串联起来。”
关注事件视觉传感器的读者应该知道，这类传感器此前虽然逐步在部分工业领域有应用，但其应用范围始终称不上广泛，尤其一度被看好的消费类应用，也少见事件视觉传感器身影。而如果将它和传统CIS图像传感器融合，是否真能创造出视觉感知的新时代？藉由与邓坚的此番对谈，从下面5个问题的角度，或许能得到更清晰的答案。
HVS融合视觉传感器，具体是什么定位？
首先要明确HVS融合视觉传感器的角色定位，如前所述它融合的是事件传感器与传统图像传感器。有关事件视觉传感器的基本原理，电子工程专辑的历史文章已经有过阐述，本文不再多做延伸——基于其“仿生”原理，事件视觉传感器的优势包括了：产生数据量少，响应速度快（体现在信息流的连续捕获，达成超高帧率），高动态范围（≥120dB）等。
从锐思智芯官网的介绍来看，HVS融合视觉技术由其原创，是一种像素级的事件感知技术与图像传感技术融合，在同一像素阵列中集成图像采样链路与事件检测电路，并通过像素内感存算一体结构设计，采用iampCOMB, GESP或IN-PULSE DiADC架构与时分复用读出机制（PixMUX架构），支持图像模式、事件模式、融合模式这三种工作形态。
图2： 融合视觉技术将APS (Active Pixel Sensing，有源像素传感或图像传感技术)与EVS (Event-based Vision Sensing，事件感知技术)两种异质路线贯穿整个技术栈——从传感器架构、像素设计到数据处理及应用级部署。
邓坚在采访中表示虽然无法透露具体的像素结构，但HVS融合视觉传感器首先是作为图像传感器存在的，应当将其视作在图像传感器上，集成事件感知能力。所以HVS传感器的定位明确为传统CIS图像传感器的迭代产品；在部分应用场景里，HVS传感器的作用也不像纯粹的事件视觉传感器那样，仅限于辅助感知，而是参与成像或者作为机器视觉的主要信息来源。
邓坚认为，HVS传感器与事件视觉传感器应用场景不同，二者不在同一赛道。“纯事件传感器可以做到10000 fps甚至更高的帧率，以更精细的方式，用在对速度有极致要求、成本不敏感的应用中。”
另一方面，相较事件视觉传感器，HVS融合传感器可输出图像。同时，在技术持续迈进以及与应用场景适配的过程里，凭借更小的像素尺寸，和在帧率、动态范围、数据量之间的权衡（如邓坚提到，HVS传感器虽不及事件传感器那么追求单点性能极致，但等效帧率和动态范围也做到了1000 fps和120dB），锐思智芯的HVS传感器也具备更强的成本优势。
由于AI对多模态信息有需求，“通常需要将多个传感器放在同一个系统内”：以手机为例，ToF、光谱传感器辅助摄像头成像即是如此。邓坚表示，随着技术的发展，这些辅助感知的传感器未来有机会与图像传感器集成在一起，成为传感器的一部分。“我们的布局也是如此，事件视觉感知将来也会融合到手机的主摄或超广角摄像头之中。这是趋势。”
图3：一款专为智能手机设计的 HVS 传感器，通过引入真实的高频运动数据，实现高质量、功耗友好的影像应用，提升移动端 AI 感知的计算效率。
在锐思智芯看来，HVS传感器未来有较大概率成为主流的视觉传感器。“尤其机器视觉感知需要真实的运动数据，除了事件感知之外，现在还没有哪种传感器能够获取这类数据；但运动数据一定需要在空间坐标内，和图像做mapping或配准。而HVS在这方面有着天然的优势，在满足应用需求方面，会有很棒的发展前景。”
能用来做什么，应用场景如何？
对于这个问题，理论上事件视觉感知的大部分应用，也适配HVS传感器——我们以往的文章也多有提及；且基于后者更低的成本+可输出图像的特性，应用场景理论上也更为广泛。这里举两个例子。
比如在汽车上，高帧率、高动态范围、低功耗、低数据量特性令事件视觉传感器适配各类场景。作为一种适用于运动快速感知的传感器，它能用于场景变化的预判：“传统图像传感器本身提供细节画面——曝光需要时间，细节信息丰富也意味着关键信息提取需要时间，也就无法做到快速响应。”
图4：在车辆逆光驶出隧道等快速变化场景，EVS的动态范围和数据量优势，可快速探测到前方障碍物，减少制动系统时间，提升主动安全
“而事件感知能够快速提供粗略的预判信息”，比如它能告诉系统附近有快速移动的物体，可能存在危险；尔后藉由图像传感器或其他传感器，去精细感知具体情况并做综合判断，“这对快速响应、提升效率、降低功耗都有价值。这一特性也同样适用于智能汽车、机器人等场景。
再比如在手机上，事件感知能力决定了HVS传感器在暗光环境下仍然能够以高带宽捕获场景的运动特征，并基于此做图像优化，实现成像的去运动模糊。而在拍摄视频时，要对视频应用后期或AI，帧和帧之间需要保持较高的场景一致性，从而对模型、算力都会有要求，事件感知数据或运动信息能够为之提供真实且经济的数据参考。
邓坚告诉我们，锐思智芯有一款小尺寸的HVS传感器，其提供的图像+运动信息可直接与主摄做mapping与“对齐”，从而绕开事件与图像需要匹配的难题——“我们和手机厂进行了不少沟通才做了这样的产品形态”。
当然，如前所述，在锐思智芯畅想的未来，HVS终将与摄像头融为一体，无论是主摄还是辅摄。而且方案是参与成像前（pre-capture）、成像中（capture）、成像后（post-capture）整个流程：
1、成像前参与测光、对焦；
2、成像中通过事件信息去记录曝光过程中的运动、变化信息；
3、成像后，可基于成像中获取的信息，显著提升帧和帧之间的一致性，参与去模糊、HDR成像、插帧等处理过程。
2023年，锐思智芯的首颗量产版HVS传感器APX003（ALPIX-Eiger）问世，目前已经落地的应用场景包括两个大方向。其一是需要捕获运动信息、具备快速感知能力的应用，如快速移动的机器人、运动相机等；其二是在降低帧率过后，作为“高度蒸馏过信息、提取过特征的低功耗传感器”用于端侧智能，如可穿戴视觉设备、手机前摄、智能家居、交通安防等。
在应用的问题上，尤为值得一提的是，邓坚在采访中多次提到HVS融合视觉传感器是AI传感器——这也是在我们看来HVS传感器最具吸引力的部分。AI要理解现实世界，基于传统图像传感器获取的高帧视频数据，由于大量冗余信息和高数据量，特征提取算力要求高、算法难度大。“比如自动驾驶，首先要用海量数据做训练，其中大部分还是无效数据，然后在其中做有效信息的提取。”
而事件视觉传感器“相当于在传感器层面就做了数据蒸馏，滤除了无效信息，存储与计算消耗都大幅下降”，一旦事件感知技术本身成熟，且生态建立起来，“一定会催生更多新兴的AI应用场景。在我们看来，它也因此对物理AI、具身智能理解物理世界有着莫大价值。”
图5：HVS融合传感器APX014（ALPIX -Pizol）在移动中实时捕捉街景：事件（左）/ 图像（右）。APX014采用全局快门和单色图像模式，提供130万像素APS/EVS分辨率
为什么要将事件与图像感知做“像素级”融合？
上述应用场景是能够展现HVS与事件视觉传感器的前景的，尤其在为AI模型提前做“数据蒸馏”这一点上。不过实际上，做一个松耦合的系统，如前所述，多模态数据获取基于不同的传感器，似乎也行。那么HVS融合视觉传感器，作为一种在“像素级”做到了事件与图像紧耦合的系统，相比一个系统用两颗不同传感器的方案，又有什么价值呢？
我们目前尚不清楚锐思智芯是如何在同一个传感器上，实现事件与图像感知像素级融合的。关注电子工程专辑此前对事件视觉传感器技术报道的读者应该清楚，其像素结构与传统CIS图像传感器的像素结构差异甚大。
采访中邓坚也提及，事件视觉传感器相当于对光电信号做微分、提取变化（对光强做对数、检测对数亮度变化），再做处理（与阈值比较并触发事件）；图像传感器则是通过积分的方式再做数据处理；这是二者很难直接融合的原因，其架构并不兼容。
不过邓坚提到HVS传感器首先作为图像传感器，和事件传感器本质上完全不同。另外，HVS传感器与传统图像传感器更为近亲还体现在，锐思智芯将HVS传感器的像素做到了更小的尺寸——“事件传感器的像素结构决定了很难把像素做小”，比如前两年的IMX636单像素尺寸为4.86μm，“我们已经能够把HVS传感器的像素做到更小——目前则已经有1.89μm的HVS传感器量产。”
锐思智芯的路线图上已经规划了一款更高分辨率的HVS传感器。“尽管技术上可以做到更多像素同时输入事件和图像，但这对事件感知而言没有实际意义”。锐思智芯官网列出的HVS传感器规格表也将分辨率拆分成APS和EVS做标注，其中EVS分辨率即为用于事件感知的分辨率或像素量。
因为一方面，运动信息并不需要那么精细；另一方面过高的事件像素也产生了数据量的压力；传感器的成本也会变得非常高。具体需要多少像素量，还是和应用场景、算法需求有关。“对不同应用，现在行业已经有比较好的事件与图像像素的比例参考。”邓坚介绍说。
回到两者“像素级”融合的价值问题：关键在于“在一颗传感器的同像素内产生事件与图像，时间和空间坐标是天然配准的——实现了原生融合”，或者说多模态信息的对齐。
异质的图像数据和事件数据做像素级对齐的难度很大：数据不同源、操作不同频；一个是精细数据，一个是运动特征；二者后端配准对算力也有要求。真正要落地就会有问题。“这也是过去有企业尝试面向手机推纯事件视觉传感器并不成功的原因。”
邓坚也谈到，纯事件视觉传感器的数据形态，和现在主流矩阵运算的数据通路不适配，要从底层打通、开发周期很长——需要改变算法体系、数据结构，对开发者而言这里面变化的代价还是很大的，比如手机的两个摄像头去做跨模态数据配准，尽管当然可以通过大模型来完成，但投入产出显然不成正比。
何时大规模应用与上量？
这番解释，实际上也明确了为什么前两年活跃于市的事件视觉传感器企业，却未能在市场上斩获理想份额。另外邓坚也总结了相关事件视觉传感器的市场经验教训：当事件视觉传感器作为产品推向市场时，面临各种具体问题，如前文提到与手机主摄的配准、不同形态数据的处理等，从产品角度看，当时事件传感器供应商也没有真正和手机厂做充分沟通，做好产品定义，等等。
但他也肯定了事件视觉传感器企业过往所做的努力，无论是传感器本身还是配套算法、生态构建，都是教育市场、加强认知的过程。“就像ToF传感器，从兴起到进入手机供应链也经历了5、6年时间，需要持续投入和积累。”“这还涉及到技术成熟度、投入产出比的问题，尤其消费电子产品在这方面尤其敏感。这是需要过程的。”
在邓坚看来，现在对HVS融合视觉感知和锐思智芯而言，是个不错的发展时机。一方面，如前文所述，更多场景对运动感知以及低功耗有更高的要求，应用端走到了对多维数据获取有切实需求的时刻——尤其AI发展起来，“从文本模型走向世界模型”，“才能以更有效的方式看到真实世界”。
图6：HVS融合传感器APX014 (ALPIX -Pizol) 实时捕捉行人运动：事件(左)/ 图像(右)。APX014采用全局快门和单色图像模式，提供130万像素APS/EVS分辨率
另一方面，事件与图像感知融合技术走向成熟，能够以更合理的投入产出比有效地解决问题；乃至契合AI发展趋势，为AI理解世界提前做好数据蒸馏，提升AI及各类应用感知世界的效率——“我们从传感器的角度已经对数据做了过滤，提供运动或事件特征，非常适合端侧小尺寸、高性价比设备的开发”。
图7：锐思智芯APX系列芯片/模组（所有产品为真实影像，图片背景由AI生成）
目前，锐思智芯的HVS融合视觉传感器已经在可穿戴、AIoT、机器人、智慧交通、安防等领域出货，“今明年（2026、2027）我们预计会是个爆发期。我们是目前这类传感器唯一实现了技术量产的公司，未来两年会有相当好的发展潜力。”
“我们的芯片技术本身已经成熟，也能够量产；中间也经历了长时间的算法和系统开发；后面铺开的速度会很快。”“今明两年预计会看到HVS传感器在更多日常产品中大量应用，毕竟它也符合AI发展趋势。”
邓坚表示，由于HVS传感器的制造没有采用任何特殊工艺，而基于图像传感器的标准工艺，国内在工艺方面也相当成熟，故而在面对未来可能爆发的市场之时，产能不会成为问题。
技术与开发生态成熟度如何？
2025年4月，锐思智芯宣布完成B轮融资，投资方包括智慧互联产业基金、浦耀信晔、中车时代投资等多家知名机构。在此之前，锐思智芯已经完成了4轮融资，资方囊括联想、OPPO、舜宇、歌尔、虹软，以及中科创星、清科创投、耀途资本等，其中有一些是连续多轮跟投，可见资本市场对HVS技术热情不减。
自2019年锐思智芯起步，至2023年HVS传感器产品问世，到现在邓坚判断很快HVS传感器面临市场爆发，期间攻坚大量工程技术问题。不少公司都在做类似技术的尝试，部分巨头也有这个方向的战略规划，因为HVS的确有很好的应用前景——只不过他们都还没有让技术和产品进入到量产阶段。邓坚告诉我们，正是基于先发优势和努力，锐思智芯率先量产了HVS传感器产品。
“在股东、Fab厂的支持下，锐思在过去6年时间里，花费大量精力，不仅让传感器在图像性能这块能够满足手机需要，还让像素加入事件感知能力，传感器也在分辨率和像素尺寸上逐步升级，适配摄像头的机械结构，同时保证低成本；而且面对两种不同频的机制，要克服信号干扰问题，从工艺、隔离、时序、芯片架构，到数据结构和算法等角度去系统性解决各种难题。”
传感器准备就绪之后，落地应用实际还涉及到系统、软件相关工作。邓坚表示，锐思智芯经过多年发展，各个链条都相对齐全，“团队里面有很多做算法、系统的同事，给到客户的不仅仅是一颗芯片，还有底层算法与方案。”
除模组、开发板、参考设计等系统级配套方案，邓坚特别提到了算法开发，首先确立底层核心算子——基于2bit事件感知数据去做算子和优化——这部分需要硬件支撑，是长期方向；其次是驱动层的算法——比如人形检测、运动检测、手势识别等相对通用的模型，基于事件特性训练和优化，形成面向客户的SDK；最后是应用层，基于SDK，配合客户或方案商去做深度定制开发。
同时，锐思智芯也提供事件数据集，期望从数据的角度来完善生态；应用开发者也更有机会基于特定场景和需求做开发。在算法与数据两个维度上的努力，最终“让大家能够集中于应用层开发”。从应用开发的角度来说，“只要发现事件或融合视觉技术的优越性，融入到应用之中，开发的难度不会很大。”
此外，锐思智芯和许多研究机构也有合作：在事件传感器本身作为视觉领域科研大热门之外，多模态信息获取越来越成为新需求。“我们和国内外的多所顶尖学校和研究机构合作，取得了一些很让人兴奋的前沿开发成果。”对于顶会的赞助、举办技术挑战赛等，也都是促进生态发展的典型操作了。
邓坚表示，HVS生态建设是个长期的过程，还“需要大企业去牵引——更多人愿意参与进来，我们的算法有了用武之地，芯片也能够得到普及。”，他希望，无论是在汽车、机器人，还是在AIoT、手机、智慧生活的方方面面，“产品得到更多人的接受和认可，成为普遍的技术，成本也能更进一步地降下来，让技术和产品更好用，持续迭代，进入正向循环。”
以下为英文全文：
Reshaping the Eye of AI: How will HVS Transform AI Machine Vision in 2026?
An Interview with AlpsenTek CEO Jian Deng
By Illumi Huang, Senior Industry Analysis, EE Times
Over the past several years, event-based vision sensors (EVS) have attracted continuous attention across the semiconductor and AI industries. Unlike conventional CMOS image sensors (CIS), these bio-inspired devices detect only changes in a scene rather than continuously sampling static imagery. Each pixel operates independently, triggering events when illumination changes exceed a threshold. This architecture enables successive temporal perception, ultra-high effective frame rates reaching tens of thousands of frames per second, and exceptionally wide dynamic range.
These characteristics make EVS an appealing candidate for applications ranging from automotive safety and industrial automation to medical imaging and consumer electronics. Yet despite years of research activities and several commercial launches, adoption has remained uneven—particularly in consumer markets.
In a recent interview, AlpsenTek founder and CEO Jian Deng offered a broader system-level perspective on why EVS has struggled to scale, and how Hybrid Vision Sensors (HVS) may provide a more practical path forward.
“The AI era demands multi-dimensional, multi-modal understanding of external worlds,” Deng explained. “Take intelligent driving as an example. A robust vision system ideally needs multiple dimensions of information: the two-dimensional image plane; depth; color or spectral data; and, critically, time and motion. These five dimensions together form the foundation of modern machine vision.”
Traditional image sensors supply spatial detail. Time-of-flight and other ranging sensors add depth. Spectral sensors extend color information. Time and motion can be extracted from video streams—but at a significant cost. Conventional video relies on fixed frame rates, produces large volumes of redundant data, and requires substantial computation to extract meaningful motion cues.
“Event-based vision works differently,” Deng said. “Inspired by the human retina, it encodes only changes as sparse event signals. It complements frame-based sensors by capturing motion information directly, efficiently, and consecutively.”
Since its founding in 2019, AlpsenTek has focused on integrating these two paradigms into a single device. Its Hybrid Vision Sensor fuses event-based and frame-based sensing at the pixel level. Over time, the company plans to further integrate additional modalities, including distance measurement and spectral sensing, to create what Deng describes as a “full-stack” vision sensor for next-generation AI systems.
Against this backdrop, the interview explored five pivotal questions that define the commercial and technical trajectory of HVS.
1. What Exactly Is the Role of HVS (Hybrid Vision Sensor)?
At its core, an HVS integrates conventional imaging with event-based vision within a single sensor. While the principles of EVS have been widely covered (EVS offer several well-recognized advantages, including low data output, fast response, and high dynamic range), the defining characteristic of HVS is its positioning: it is first and foremost an image sensor, not a pure event sensor.
According to AlpsenTek, HVS implements proprietary pixel-level fusion by integrating image-sampling circuits and event-detection circuits within the same pixel array. Using in-pixel sensing and processing architectures—such as iampCOMB, GESP, and IN-PULSE DiADC—along with time-interleaved readout technologies like TILTech and PixMUX, HVS supports three operating modes: conventional image mode, event mode, and hybrid mode.
While Deng did not disclose detailed pixel architectures, he emphasized that HVS should be understood as an evolutionary step for CIS rather than a parallel technology. In many applications, it functions as an imaging sensor primary rather than a supplementary device.
This distinction is important when comparing HVS with pure EVS sensors. “Pure event sensors can exceed 10,000 frames per second,” Deng noted. “They are well suited for ultra-high speed scenarios, where cost sensitivity is lower.” HVS, by contrast, balances multiple metrics: frame rate, dynamic range, pixel size, power consumption, data volume, and cost.
In contrast to pure event-based sensors, HVS is capable of delivering high-quality image data. Through ongoing optimization, AlpsenTek’s HVS sensors can easily achieve an equivalent frame rate of1,000 fps and a dynamic range of 120 dB, while maintaining significantly lower cost than pure event sensors. This balance makes it suitable for high-volume markets.
Deng also highlighted a broader industry trend toward sensor integration. In smartphones, for example, ToF and spectral sensors already support imaging pipelines. Over time, these modalities are increasingly being absorbed directly into image sensors.
“Our roadmap follows this direction,” Deng said. “Event-based sensing will eventually be integrated into the main or ultra-wide camera. It becomes part of the sensor, not a separate component.”
From AlpsenTek’s perspective, HVS addresses a fundamental gap in data registration of vision systems. “Authentic motion information is essential,” Deng said. “There is no other sensing approach that captures motion as efficiently as EVS does. But motion data must ultimately be mapped or registered with spatial image data within temporal-spatial coordinates. That is where HVS has a natural advantage.”
2. What Can HVS Do? Application Scenarios
In principle, any application suited to event cameras can also benefit from HVS. In practice, the ability to output standard image data significantly broadens the range of viable use cases.
Automotive systems illustrate this advantage. Event-based sensing offers high dynamic range, fast response, low latency, and low data redundancy—attributes well suited to environments with rapidly changing lighting. Deng described event sensing as enabling “pre-emptive assessment.”
“Traditional image sensors deliver rich visual detail,” he said. “But exposure takes time, and extracting critical information from that detail also takes time. This lowers system response speed.”
Event-based sensing, by contrast, can quickly provide coarse input data for vehicles, such as the presence of fast-moving objects. These cues then trigger more detailed semantic analysis through sophisticated algorithms, reducing response time, improving safety, and lowering power consumption.
In smartphones, HVS supports motion capture with high temporal resolution, even under low-light conditions. Motion information can significantly improve image enhancement results such as deblurring, HDR imaging, and video frame interpolation. For video capture and AI-based post-processing (AI editing), it provides a sparse and consecutive event-based reference that maintains temporal consistency across distinctively-gapped image frames.
According to Deng, AlpsenTek has also developed compact HVS devices designed to align more directly with smartphone main cameras, reducing the complexity of registering heterogeneous data from separate sensors. This method emerged through close collaboration with smartphone OEMs and aims to integrate HVS directly into phone camera modules.
From a system perspective, HVS contributes across the imaging pipeline. Before capture, it assists with metering and focus. During exposure, it continuously records motion and changes. After capture, it improves frame registration and supports advanced processing such as deblurring, HDR, and frame interpolation.
In 2023, AlpsenTek introduced its first mass-produced HVS, the APX003 (previously known as ALPIX-Eiger). Current deployments fall into two main categories. One targets rapid-sensing applications such as robots and action cameras. The other positions HVS as a low-power, data-distilled and feature-extracted sensor for edge AI, supporting XR devices, smartphone front cameras, smart-home devices, and traffic surveillance.
Deng repeatedly emphasized that HVS is fundamentally an AI sensor. “Training AI systems with high-frame-rate video is extremely inefficient,” he said. “Most of the data is not informative, yet it drives enormous storage and computation costs.”
Event-based sensing performs data distillation directly at the sensor, filtering out non-contributory information. As a result, it enables more efficient physical AI and embodied systems that interact with the real world in real time.
3. Why Pursue Pixel-Level Fusion?
Given the availability of multi-sensor systems, a natural question arises: why pursue tightly integrated, pixel-level fusion rather than using separate sensors?
The challenge lies in the fundamental mismatch between image and event sensor architectures. Event sensors effectively compute temporal derivatives of photoelectric signals, often operating on logarithmic light intensity and triggering event signals based on threshold changes. Image sensors, by contrast, integrate light over time to form frames.
Deng acknowledged this difficulty but stressed that HVS is designed as an image sensor with integrated event capability, not a mere juxtaposition of two distinct sensors. This design choice enables its pixel arrays comparable to CIS. AlpsenTek has demonstrated a roadmap scaling down to smaller pixel size, with a 1.89 µm HVS already in mass production. By comparison, some commercially available pure event sensors such as IMX636 use pixels larger than 4 µm.
Although AlpsenTek’s roadmap includes high-resolution HVS devices, Deng cautioned that there is no practical value for put every pixel “hybrid”. Motion information supplied by EVS are typically sparse and coarse, yet excessive event pixel counts increase data volume and cost without proportional benefit.
As a result, AlpsenTek specifies APS and EVS resolutions separately, allowing event resolution to be tailored to application needs.
The core value of pixel-level fusion, Deng concluded, is native spatiotemporal alignment. When image and event data originate from the same pixel with subtle readout circuit design, spatial and temporal registration is inherent. This eliminates the need for external synchronization and computationally expensive post-alignment.
Dual-sensor architectures face intrinsically challenges in aligning heterogeneous data operating at different sampling rates. These challenges have contributed to the limited success of earlier attempts to introduce pure event sensors into consumer devices.
“Supporting event data also requires changes in algorithm frameworks and data structures,” Deng noted. “The development burden is huge, and the return on investment is often disproportionate.”
4. When Will HVS Reach Mass-Market Volume?
The difficulties faced by earlier event-sensor suppliers offer important lessons. Standalone event sensors struggled with integration, data handling, and unclear product positioning. Many suppliers also lacked deep engagement with smartphone OEMs.
Deng acknowledged the role of early pioneers in building awareness and ecosystem foundations. He compared the trajectory to ToF sensors, which required years to move from emergence to widespread smartphone adoption.
In Deng’s view, the timing is now more favorable. On the demand side, AI applications increasingly require efficient motion perception and low power consumption. On the supply side, HVS technology has matured to deliver practical ROI.
By distilling data at the sensor level, HVS aligns well with the needs of compact edge devices. AlpsenTek’s sensors are already shipping in wearables, AIoT, smart transportation, and security applications.
“We expect 2026 and 2027 to be a period of rapid growth,” Deng said. “AlpsenTek is currently the only company that brought this type of sensor into mass production, and we have already completed extensive algorithm and system development.”
Importantly, HVS manufacturing relies on standard image-sensor process flows, rather than specialized fabrication steps. As a result, Deng does not anticipate production capacity becoming a bottleneck as demand increases.
5. How Mature Are the Technology and Development Ecosystems?
In April 2025, AlpsenTek completed Series B financing round, with investors including Wisdom Internet Industry Fund, PuXin Capital, CRRC Times Investment, and others. Prior to this, AlpsenTek had secured four rounds of funding backed by leading industry players and funds, including Lenovo, OPPO, Goertek, ArcSoft, Sunny Optical, as well as Casstar, Zero2IPO, Glory Ventures and so forth (with some participating across multiple rounds). This sustained investment reflects long-term confidence in HVS technology.
From its founding in 2019 to the launch of its first HVS products in 2023, the company addressed numerous challenges, including pixel integration, interference mitigation, process optimization, and breaking mechanical constraints for camera modules.
Beyond hardware, Deng emphasized the importance of software and systems. A significant portion of AlpsenTek’s team focuses on hybrid-oriented algorithms and system development. The company delivers not only sensors, but also development boards, reference designs, SDKs, and datasets.
Algorithm development spans three layers: hardware-optimized core operators for 2-bit event data, driver-level models and SDKs such as motion and human detection, pose and gesture recognition, and application-layer frameworks formulated by customers. AlpsenTek also provides event datasets tailored to specific scenarios to lower development barriers.
The company collaborates with academic top tier institutions globally, and sponsored top academic conference on computer vision and others. These partnerships support frontier research in multi-modal sensing and perception and ecosystem development.
“The HVS ecosystem is a long-term effort,” Deng concluded. “As more participants adopt the technology, applications mature, iteration accelerates, and the ecosystem enters a positive cycle.”
热门跟贴