Edge AI-Powered Obstacle Detection and Efficiency Analysis in Autonomous Lawn Mowers for Smart Agriculture

Abstract

Idowu Olugbenga Adewumi, Titus Adeyinka Ilori and Joseph Abayomi Adebisi

The demand for autonomous agricultural machines is increasing with the global push toward smart farming and sustainable land management. This study showcases a framework for obstacle detection and efficiency analysis in an autonomous lawn mower, utilizing Edge AI technology. The system employs a pruned MobileNetV2 convolutional neural network operating on an ESP32-CAM alongside ultrasonic sensors for instantaneous obstacle recognition, attaining 92.3% classification accuracy, a macro-averaged AUROC of 0.94, and a mean Average Precision (mAP) of 0.92, with an inference latency of 85 ms per frame (approximately 11.8 FPS). Field trials were carried out at two locations featuring 12 randomized plots (10 m × 10 m each) following a 3 × 2 × 2 factorial design (vegetation height: short/medium/tall; terrain: level/uneven; moisture: dry/wet) with five repeats per condition (totaling 120 trials).

The mower reached a theoretical maximum field capacity (FCt) of 1.23 ha/hr, an actual field capacity (FCe) of 1.05 ha/hr, and a peak field efficiency (η) of 85.4% ± 2.3% CI on short, flat, dry areas. Efficiency decreased to 71.0% ± 3.5% CI on high, irregular, damp areas. ANOVA validated substantial impacts of surface (p < 0.01, η2 = 0.21), grass height (p < 0.01, η2 = 0.18), and moisture (p < 0.05, η2 = 0.11) on efficiency. An ablation study indicated that excluding ultrasonic sensors decreased AUROC from 0.94 to 0.86, whereas enhancing image resolution from 224 × 224 to 320 × 320 px raised IoU from 0.78 to 0.82 but boosted latency from 85 ms to 124 ms. Analysis of failures showed that obstacle misclassification happened in 4% of trials, low-light inaccuracies in 6%, ultrasonic blind spots in 5%, and slippage on wet grass in 3%, all countered by built-in safety interlocks. The findings indicate that the suggested system merges real-time Edge AI perception with quantitative analysis of field efficiency, offering a scalable and secure framework for intelligent lawn mower. Future extensions will explore multi-sensor fusion, SLAM-based navigation, and autonomous docking/charging to enable fully autonomous, continuous operation in diverse outdoor environments.

PDF