Hyper-Fast
Visual Tracking
Robust Single-Object Tracking (SOT) with Re-ID capability for complex environments.
Optimized for low-power edge detection platforms. Achieving 2ms-5ms inference latency on standard embedded NPU hardware while maintaining lock-on robustness during occlusion.
Performance Data
Our algorithm is heavily optimized for INT8 quantization on RISC-V and ARM-based TPU platforms. It delivers industrial-grade real-time performance even on entry-level chips.
Test Conditions
- Input Size: 128x128 / 256x256
- Quantization: INT8
- Batch Size: 1
| Platform (Computing Power) | Inference Speed (Latency) |
|---|---|
CV184x 1.5 TOPS | 5ms |
CV186x 7.2 TOPS | 2ms Ultra-Fast |
CV181x 0.5 TOPS | 9ms |
Robust Real-World Tracking
Our proprietary algorithm maintains consistent lock-on from initial acquisition to long-range tracking, filtering out complex background clutter and static obstacles in real-time.
1. Initial Lock-On
Acquire and lock onto the target before occlusion.
2. Occlusion Handling
Maintain target identity through brief occlusion.
3. Rapid Re-Lock
Reacquire and lock immediately once the target reappears.
Ultra-Low Latency
<5ms processing time ensures direct closed-loop control for UAVs without jitter or oscillation.
Lightweight Model
Highly optimized INT8 quantization allows deployment on cost-effective chips like CV181x.
Robust Re-ID
Maintains target identity even after total occlusion, preventing "target drift" to background objects.
Multi-Scale Input
Supports 128x128 for speed and 256x256 for precision, adaptable to different sensor resolutions.
Integrate AI Tracking Today
Available as a software SDK for CV18xx series chips or pre-flashed on our AFC-V1 Flight Controller.