CASE STUDY · EMBEDDED AI
Real-Time IR Target Detection on Embedded Compute
Designing AI/ML perception for infrared imagery under strict latency, memory, and false-alarm constraints.
Problem
Infrared target detection is a systems problem disguised as a model-training problem. The detector must identify dim, difficult targets in cluttered environment, but the useful operating point is shaped by more than model accuracy. False alarms, inference latency, memory footprint, CPU utilization, tracker behavior, and the target processor all influence whether the perception system is useful in practice.
I led R&D work to move machine-learning target detection and classification from experimental models toward real-time embedded execution and system-level evaluation.
Constraints
- Limited compute and memory on the embedded target.
- A real-time processing budget rather than offline throughput.
- Dim-target imagery where preprocessing and output resolution matter.
Approach
I treated the detector as one component in an end-to-end perception chain. Model architecture work was paired with deployment analysis, inference optimization, data-pipeline improvements, and software-in-the-loop integration.
Key engineering threads included:
- Architecture search and model design to improve the accuracy-efficiency trade space.
- Embedded inference optimization to reduce latency, memory footprint, and CPU pressure.
- Flight-test imagery pipelines to make training and evaluation data easier to ingest and preprocess.
Results
The resulting CNN design improved model accuracy while also reducing inference time and memory usage, enabling real-time embedded processing. Integration into a software-in-the-loop testbed demonstrated improved model accuracy.
Engineering notes
The important lesson was that the best detector is not necessarily the model with the highest isolated benchmark score. For an embedded perception stack, the right objective is closer to a constrained systems optimization problem:
\[\text{utility} = f(\text{detection quality}, \text{latency}, \text{memory})\]That framing naturally connects model design, neural architecture search, embedded optimization, and system-level verification.