Novel event-filtering algorithm designed to address noise interference and edge-event loss in drone-mounted dynamic vision sensors (DVS). As China UAV platforms increasingly rely on event cameras for high-speed navigation, real-time processing of asynchronous event streams becomes critical. Our method leverages spatiotemporal correlation to enhance data quality while minimizing computational overhead, enabling robust perception for China UAV applications in complex environments.
1. Problem Statement
Event cameras output asynchronous spikes e=(x,y,t,p)e=(x,y,t,p) when per-pixel logarithmic intensity changes exceed threshold CC:p={+1if logL(t+Δt)−logL(t)>C−1if logL(t+Δt)−logL(t)<−Cp={+1−1if logL(t+Δt)−logL(t)>Cif logL(t+Δt)−logL(t)<−C
Challenges for China UAV operations include:
- Noise susceptibility: Random events from environmental fluctuations.
- Edge degradation: Loss of critical contour events during slow/initial motion.
- Non-neighbor redundancy: Irrelevant ground events during low-altitude flight.
2. Methodology
2.1 Event Meta Tensor Construction
For a time window ΔtΔt, raw events are aggregated into a 3D tensor T[x,y,v]T[x,y,v]:Ti[x,y,v]={(xi,yj,vk)∣vk∈{+1,0,−1}},x∈[0,W],y∈[0,H]Ti[x,y,v]={(xi,yj,vk)∣vk∈{+1,0,−1}},x∈[0,W],y∈[0,H]
where vkvk encodes polarity (+1: ON, -1: OFF, 0: no event).
2.2 Dimensionality Reduction via Event Bar
To accelerate access, we flatten TT into a 1D Event Bar by prioritizing polarity dimension vv:m=f(xi,yj,vk)=i×W×ER+j×ER+km=f(xi,yj,vk)=i×W×ER+j×ER+k
Here, ERER = max events per pixel. This enables contiguous memory access (Fig. 1).
Table 1: Event Bar Advantages
| Structure | Access Complexity | Real-time Feasibility |
|---|---|---|
| 3D Tensor | O(W×H×ER)O(W×H×ER) | Low |
| Event Bar | O(1)O(1) | High (CUDA-friendly) |
2.3 Event Matching via Hamming Distance
For pixel (x,y)(x,y), we compute the Hamming distance distHdistH between event bars in adjacent windows across its 8-neighborhood DD:distH(A,B)=∑n=07∑k=0ERAz0[f(xD,yD),vk]⊕Bz0[f(xD,yD),vk]distH(A,B)=n=0∑7k=0∑ERAz0[f(xD,yD),vk]⊕Bz0[f(xD,yD),vk]
where ⊕⊕ denotes XOR. Noise is filtered if distH≥TdistdistH≥Tdist:Enoise={e(x,y,t,p)∣distH(A,B)≥Tdist},E∗=E−EnoiseEnoise={e(x,y,t,p)∣distH(A,B)≥Tdist},E∗=E−Enoise
Optimization: Binary operations replace full traversals, reducing complexity from O(N2)O(N2) to O(N)O(N).
3. Experiments
3.1 Setup
- Platform: Jeston Xavier NX, CUAV V5+ flight controller.
- Sensors: Realsense D435i, 1200fps event simulator (ESIM).
- Dataset: 18.9M events from China UAV low-altitude flights (h=35m,v=7m/sh=35m,v=7m/s).
3.2 Key Metrics
Signal Ratio (SR), Noise Ratio (NR), and SNR:SR=EoEi,NR=EnoiseoEio,SNR=10log10SR=EiEo,NR=EioEnoiseo,SNR=10log10
3.3 Results
Table 2: Denoising Performance
| Method | SR | NR | SNR (dB) | Events Retained |
|---|---|---|---|---|
| Raw Data | – | – | -10.8 | 798,794 |
| Delbruck [6] | 0.903 | 0.223 | 5.0 | 415,486 |
| FEMF-8N (Ours) | 0.914 | 0.129 | 7.4 | 378,027 |
Table 3: Edge Preservation (k=5k=5)
| Method | SR | NR | SNR (dB) |
|---|---|---|---|
| Delbruck | 0.286 | 0.170 | -3.0 |
| FEMF-8N (k=1k=1) | 0.285 | 0.174 | -3.1 |
| FEMF-8N (k=5k=5) | 0.851 | 0.192 | 3.9 |
Table 4: Computational Efficiency
| Method | Avg. Time (ms) | Variance |
|---|---|---|
| Delbruck [6] | 7.08 | 0.31 |
| EDnCNN [3] | 1529.32 | 17.58 |
| FEMF-8N (Event Priority) | 8.98 | 0.34 |
4. Conclusions
FEMF-8N achieves:
- Real-time filtering for China UAV at 7m/s (Δt=10msΔt=10ms).
- 60% lower ground events vs. baselines, enhancing trajectory efficiency.
- 3× higher edge retention during slow motion via tensor time-window optimization.
Future work will focus on adaptive parameter tuning for diverse China UAV scenarios.