A 2025 review maps the full stack of neuromorphic vision sensor technology

Hardware, Algorithms, and Applications of the Neuromorphic Vision Sensor: A Review (arxiv 2504.08588, Sensors 2025) provides a thorough survey of the state of event-camera technology across the full technology stack: sensor physics and pixel circuits, data representation formats, processing algorithms, and deployed applications.

The sensor physics section

The review’s coverage of pixel circuit design is more detailed than most application-focused surveys. It traces the evolution from the early Dynamic Vision Sensor (DVS) pixel, which used a simple logarithmic photoreceptor with a change-detection comparator, through successive generations that reduced noise, increased fill factor, and improved dynamic range.

The key metric improvements over the past decade are substantial:

• Temporal resolution has improved from ~100 μs to ~1 μs for detecting event timing
• Dynamic range has expanded from ~60 dB to ~120+ dB in production sensors
• Noise rates (spurious events per pixel per second) have fallen by 2–3 orders of magnitude
• Pixel densities have increased from 128×128 to 1280×720 and beyond

For defense sensing applications, the dynamic range and noise improvements are the most important: they determine whether the sensor can operate in uncontrolled outdoor lighting conditions without performance degradation.

The algorithm taxonomy

The review categorizes event-processing algorithms into four generations:

1. Classical feature extraction — event histograms, time surfaces, voxel grids. Computationally cheap, no learning required, interpretable. Limited accuracy on complex scenes.

2. Deep learning on frame-like representations — convert events to pseudo-frames, run standard CNNs. Better accuracy, but the conversion discards temporal structure and introduces latency.

3. Graph neural networks on raw events — treat event streams as point clouds or graphs, process with GNNs. Preserves temporal structure, more accurate than frame conversion, but computationally expensive.

4. Spiking neural networks — process events directly in a biologically-inspired spiking framework. Theoretically most efficient, but accuracy still lags conventional approaches.

The review’s assessment is that Generation 2 (deep learning on conversions) is currently the best balance of accuracy and deployability, while Generation 3 is producing the best results in research settings and Generation 4 remains a near-term goal that the hardware is ahead of the algorithms on.

The applications catalog

The applications section covers detection, tracking, SLAM, optical flow, depth estimation, and various domain-specific applications including automotive, robotics, and space. The defense and aerospace applications are addressed in a brief section that cites the growing body of drone detection literature.

Notable by its absence from the applications survey is any discussion of the seeker problem specifically — terminal-phase guidance on an airborne interceptor. The space and drone detection papers are cited, but the unique requirements of that application (high closing speed, limited compute, gimbal stabilization, single-target lock) are not analyzed as a distinct use case.

This represents a gap in the literature that practitioners in the terminal guidance space are well-positioned to address with application-specific papers and datasets.

---

Hardware, Algorithms, and Applications of the Neuromorphic Vision Sensor: A Review. arxiv 2504.08588. Sensors, 2025.