Validation study demonstrates efficient fire and smoke detection on resource-constrained FPGA hardware, indicating viable low-power edge surveillance for early wildfire prevention.
Key Points
To develop and evaluate YOLO-FSD, a lightweight target detection network derived from YOLOv4-tiny tailored for deployment on resource-constrained edge FPGA devices.
Engineered YOLO-FSD by incorporating depthwise separable convolutions, inverted residuals, a lightweight semantic enhancement block, a lightweight P4 head, and a shallow detail compensation branch.
Applied batch normalization fusion and 16-bit integer quantization to perform fixed-point forward inference on a Xilinx Zynq-7020 FPGA platform.
Benchmarked detection accuracy on a combined test set from the D-Fire and New Fire and Smoke datasets against baseline YOLOv4-tiny.
YOLO-FSD achieved a 69.36% mAP50 with 3.951 M parameters and 1.520 G MACs, improving mAP50 by 2.76 percentage points while reducing parameters by 32.76% and MACs by 55.53% compared to YOLOv4-tiny.
Direct 16-bit fixed-point inference on the Zynq-7020 FPGA attained 69.14% mAP50, representing a negligible 0.22 percentage point loss compared to the 32-bit floating-point baseline.