Dynamic and Active-Pixel Vision Sensor (DAVIS)

• DAVIS is a hybrid vision sensor that integrates synchronous global-shutter APS frames with asynchronous DVS events for dynamic scene capture.
• It supports diverse operational modes, using APS frames for absolute brightness and event data for high-speed, low-latency processing in robotics and driving applications.
• Its design offers advantages such as reduced motion blur, improved dynamic range (up to 140 dB), and co-registered outputs that enhance algorithms for SLAM, visual odometry, and tracking.

Dynamic and Active-pixel Vision Sensor (DAVIS) denotes a hybrid vision sensor that integrates a conventional global-shutter active-pixel sensor (APS) and an event-based vision sensor in the same pixel array, so that the device outputs synchronous grayscale frames together with asynchronous brightness-change events. Because the two modalities are generated by the same physical pixels, the frame stream and the event stream are intrinsically co-registered, which makes DAVIS a central platform for algorithms that combine classical image processing with event-based sensing in robotics, visual odometry, SLAM, high-speed tracking, and related domains (Mueggler et al., 2016, Mohamed et al., 2020).

1. Hybrid architecture and sensor family

DAVIS is characterized in the literature as a single-sensor fusion architecture: a conventional APS path provides standard grayscale images carrying absolute brightness, while a Dynamic Vision Sensor (DVS) path reports only brightness changes and does so asynchronously. A defining hardware property is that both outputs are produced by the same pixel lattice, so there is no spatial offset between the image stream and the event stream. This differs from multi-camera rigs in which one must estimate and maintain extrinsic calibration between separate frame and event sensors (Mueggler et al., 2016).

Representative devices in the literature include the DAVIS240 family at $240\times 180$ resolution and the DAVIS346B prototype at $346\times 260$. The DAVIS346B is described as having 2.1X more pixels than an earlier DAVIS design, on-chip column parallel ADCs, and APS output up to $50\,\text{fps}$; in driving recordings its APS rate varied between $10\,\text{fps}$ and $50\,\text{fps}$ depending on exposure duration. The DAVIS240 line appears in tracking, optical-flow, grasping, and dataset work, while the DAVIS346 family also appears in driving and pixel-model studies (Binas et al., 2017, Stoffregen et al., 2018, Graca et al., 12 May 2025).

This architecture is significant because it permits several modes of operation within one sensor. APS frames support initialization, visualization, feature extraction, and compatibility with conventional computer-vision pipelines, whereas DVS events support low-latency, high-temporal-resolution processing when scene dynamics are rapid or illumination is difficult. This duality is the core reason DAVIS is repeatedly positioned as a bridge technology between frame-based and event-based vision (Mueggler et al., 2016, Guo et al., 2019).

2. Event generation, output semantics, and operating characteristics

In DAVIS-based work, an event is typically represented as $e=\langle x,y,t,p\rangle$ or $e_k=(x_k,y_k,t_k,p_k)$, where $x,y$ are pixel coordinates, $t$ is a timestamp, and $p\in\{\pm 1\}$ is the polarity indicating whether brightness increased or decreased. One formulation writes the trigger condition as

$346\times 260$0

with $346\times 260$1 a contrast threshold; related simulator formulations describe events as threshold crossings in log intensity and reconstruct log intensity by accumulating signed increments of size $346\times 260$2 from an initial image (Guo et al., 2019, Mueggler et al., 2016).

The event output is asynchronous and pixel independent. In the slip-prevention work, DAVIS is explicitly described as using level-crossing sampling and Address Event Representation (AER), with asynchronous transmission through a request/acknowledge protocol. In this mode, only pixels whose incident light changes sufficiently relative to a reference generate data, which avoids processing unchanged pixels and reduces redundant output relative to conventional frame cameras (Masoumian et al., 2021).

Across the cited literature, DAVIS and event cameras are associated with microsecond-scale latency, microsecond timestamp resolution, high pixel bandwidth, and very high dynamic range. Reported dynamic-range figures include about $346\times 260$3 and $346\times 260$4, contrasted with roughly $346\times 260$5 for standard cameras, and one paper states that event output can reach up to 10 million events per second in high dynamic environments. In driving-oriented descriptions, DVS events are said to have dynamic range $346\times 260$6, effective frame rates $346\times 260$7, and data rates comparable to $346\times 260$8 image sensors (Mueggler et al., 2016, Mohamed et al., 2020, Binas et al., 2017).

These properties explain why DAVIS is repeatedly used in settings where conventional fixed-rate cameras suffer from motion blur, saturation, or excessive latency. The event stream is not a dense image sequence; it is a sparse, change-driven signal whose bandwidth depends on scene dynamics. APS frames remain important because they provide absolute intensity structure that events, by definition, do not (Mueggler et al., 2016, Guo et al., 2019).

3. Algorithmic use patterns: hybrid fusion, asynchronous inference, and event-native estimation

A recurrent DAVIS design pattern is a division of labor between APS and DVS. In the corner-tracking work, grayscale APS frames are treated as keyframes on which Harris corners are extracted, while the DVS stream asynchronously propagates those corners between frames. The method validates event-corners in a $346\times 260$9 local surface of active events, tracks them in a $50\,\text{fps}$0 neighborhood, estimates local plane geometry in $50\,\text{fps}$1 with a randomized Hough transform, and updates corner locations up to 100 times during the blind time between consecutive intensity images (Mohamed et al., 2020).

A second algorithmic pattern is explicit rejection of frame-like accumulation as the primary computational abstraction. In the steering-angle paper, prior DAVIS/event methods are criticized for accumulating events over windows such as $50\,\text{fps}$2 and then processing them as ordinary images. The proposed alternative constructs timestamp-specific event matrices $50\,\text{fps}$3, encodes them into vectors $50\,\text{fps}$4, integrates them with a GRU, and uses the resulting asynchronous representation to generate attention over APS features before final regression. The method is reported as best in three of four driving scenarios, which the authors interpret as evidence that preserving the asynchronous character of events is beneficial for motion estimation (Guo et al., 2019).

A third pattern is event-native motion estimation from slices or local spatiotemporal structure. ABMOF uses adaptive time slices of accumulated DVS events, block matching with SAD, multi-scale pyramids, and feedback control that tries to keep average match distance near $50\,\text{fps}$5. Its slice-formation policy includes ConstantDuration, ConstantEventNumber, and AreaEventNumber rotation rules, with the last designed to react to local event hotspots rather than global activity. Although the method operates on DVS events rather than APS frames, it was evaluated on an advanced DAVIS with $50\,\text{fps}$6 resolution, and the paper positions it as especially suitable for compact logic circuits and high-speed embedded use (Liu et al., 2018).

Taken together, these works show that DAVIS does not prescribe a single computational model. It supports at least three distinct regimes: APS-initialized event refinement, fully asynchronous neural processing, and event-native correspondence or flow estimation. This suggests that the sensor’s practical identity is as much algorithmic as hardware-defined.

4. Robotic manipulation, driving, and high-speed reactive perception

In robotic manipulation, DAVIS has been repurposed as a tactile sensor by mounting the camera behind a transparent acrylic-and-silicone gripper so that contact deformation becomes optically observable. In the Baxter slip-prevention system, contact is inferred from negative-polarity events generated when an object blocks light at the silicone interface; slip is associated with positive-polarity activity in the normalized contact area. The controller samples every $50\,\text{fps}$7, uses the number of positive-polarity events as the control error, and employs a PI controller to reduce the gripper-position value when slip begins. The system was tested in 200 trials on four objects with different shapes, sizes, weights, and materials, with qualitative validation from an FSR402 force-sensitive resistor (Masoumian et al., 2021).

In autonomous-driving research, DAVIS has been used both as a data source and as an input modality for end-to-end learning. The DDD17 work introduced over $50\,\text{fps}$8 of annotated DAVIS driving recordings at $50\,\text{fps}$9 resolution, covering highway and city driving in daytime, evening, night, dry, and wet weather, together with vehicle telemetry such as steering angle, speed, GPS, throttle, and brake. The same line of work motivated using APS and DVS as complementary streams: APS provides conventional scene appearance, while DVS contributes low-latency temporal contrast information under difficult illumination and fast motion (Binas et al., 2017).

DAVIS has also supported closed-loop mobile-robot behavior. In the predator-prey system, APS frames were produced at a fixed $10\,\text{fps}$0, while DVS histograms of fixed event count were produced at a variable rate that depended on robot speed and scene activity. The selected runtime system used $10\,\text{fps}$1 grayscale inputs and achieved a mean $10\,\text{fps}$2 error in angular estimation of prey direction from the combined ten-class CNN output. The practical point was not merely recognition accuracy, but an automatically varying latency-power tradeoff: more computations occurred when image dynamics increased (Moeys et al., 2018).

High-speed reactive perception is a related, though not always explicitly DAVIS-specific, domain. A DVS-based spatiotemporal network for time-to-collision and impact-location prediction was designed around the same event representation that DAVIS emits. It used 10 causal exponential filters across time scales from $10\,\text{fps}$3 to $10\,\text{fps}$4, produced predictions every $10\,\text{fps}$5 for darts, and reported performance on objects moving up to about $10\,\text{fps}$6. A plausible implication is that DAVIS can serve not only as a hybrid sensor for scene understanding but also as the event source for reflex-like control loops (Bisulco et al., 2020).

5. Datasets, benchmarking infrastructure, and simulation

DAVIS became especially influential through benchmarking infrastructure. The event-camera dataset and simulator paper released real and synthetic sequences for pose estimation, visual odometry, and SLAM, all centered on a DAVIS at $10\,\text{fps}$7 resolution. The released data include asynchronous events, global-shutter intensity frames at about $10\,\text{fps}$8, inertial measurements at $10\,\text{fps}$9, and ground-truth camera poses from motion capture at $50\,\text{fps}$0 for indoor sequences. Data were released both as standard text files and as binary rosbag files, with explicit documentation for events.txt, images.txt, imu.txt, groundtruth.txt, and calib.txt (Mueggler et al., 2016).

The same ecosystem established calibration and synchronization practices that became standard reference points for DAVIS work. The dataset paper describes intrinsic calibration with a pinhole plus radial-tangential model, hand-eye calibration so that released poses correspond to the event camera itself, and a camera-to-IMU transform estimated with Kalibr. It also records caveats: no hardware synchronization between motion capture and DAVIS clocks, clock drift of about $50\,\text{fps}$1, and a visual-stream lag of about $50\,\text{fps}$2 in the IMU due to low-pass filtering (Mueggler et al., 2016).

DDD17 extended this role to end-to-end driving. It stored DAVIS APS and DVS data together with vehicle telemetry in HDF5, used host-computer timestamps for synchronization, and released tools such as view.py and export.py for visualization and frame export. The dataset was explicitly framed as the first open dataset of annotated DAVIS driving recordings, and its lasting importance lies less in a single benchmark score than in standardizing how APS+DVS driving data can be recorded, indexed, and supervised (Binas et al., 2017).

Simulation has also evolved from phenomenological event generation toward circuit-derived pixel modeling. A recent DVS pixel model, validated against measurements from a DAVIS346 test pixel, derives large-signal differential equations from the photoreceptor and source-follower circuitry, fits parameters from SPICE and pixel measurements, and uses first-passage-time theory for stochastic event generation. The reported benefit is accurate noise prediction with timesteps more than $50\,\text{fps}$3 times larger than would otherwise be required and wall-clock speedups of over $50\,\text{fps}$4. Its stated scope is the DVS half of DAVIS rather than a full APS+DVS joint simulator (Graca et al., 12 May 2025).

6. Limitations, misconceptions, and adjacent directions

A common misconception is that DAVIS implies the obsolescence of frame-based sensing. The literature instead repeatedly assigns complementary roles to APS and DVS. APS frames remain important for absolute intensity, calibration, visualization, stable corner extraction, and final prediction heads; several successful methods, including asynchronous steering-angle estimation and hybrid corner tracking, still depend on APS features even while arguing for event-native temporal processing (Mohamed et al., 2020, Guo et al., 2019).

A second misconception is that raw events directly provide full texture, color, or HDR images. Multiple papers treat this as precisely the unresolved problem. Texture reconstruction work states that DVS needs extra information apart from output spikes for natural-scene visualization; color reconstruction with a DVS requires active colored illumination and a static scene; event-based HDR imaging in AsynHDR requires LCD-based temporal incident-light modulation and reconstructs brightness from event timing rather than from passive event output alone. The AsynHDR system reports about $50\,\text{fps}$5 dynamic range, but it also states that output frame rate is limited to $50\,\text{fps}$6 and that fast motion is problematic because radiance is encoded over time (Zhu et al., 2019, Cohen et al., 2022, Wu et al., 2024).

A third limitation is application dependence. In tactile grasping, performance depends on the optical-mechanical transduction layer: silicone that is too transparent increases background noise, too thick blurs contact, and too thin weakens deformation cues; larger contact areas are easier than spherical contacts; robot-arm vibration perturbs force readings. In event-based optical flow, slice quality is decisive, and the ABMOF controller can oscillate because it uses a bang-bang update of $50\,\text{fps}$7 around the target $50\,\text{fps}$8 (Masoumian et al., 2021, Liu et al., 2018).

These constraints do not diminish DAVIS’s importance; they define it more precisely. DAVIS is best understood not as a universal replacement for frame cameras or pure event cameras, but as a calibrated hybrid platform whose value emerges when algorithms exploit the asymmetry between synchronous APS images and asynchronous DVS events rather than forcing either modality to imitate the other. A plausible implication is that future DAVIS research will continue to be hybrid in two senses at once: hybrid in sensing hardware, and hybrid in the algorithmic coupling of absolute appearance with sparse temporal change.