Parallel Execution of the Viola-Jones Algorithm on MCUs for Low-Cost Automated Pest Detection

This paper presents an energy-optimized solution for on-device pest detection, preventing image streaming from the traps installed in the monitored orchard surface. We optimize the Viola-Jones object detector to run on a low-cost 8-core MicroController (MCU), which is coupled to a low-power image sensor. Thanks to the adopted memory management scheme and the algorithm parallelization, our solution presents a processing latency of only 0.4 sec per image while scoring >80% detection accuracy on low-quality images. This design reduces the memory requirements and complexity of precedent processing pipelines based on Deep Neural Networks and presents a power envelope > 10x lower than more expensive existing solutions.