Anomaly Detection for Agricultural Vehicles Using Autoencoders.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Safe operation of autonomous agricultural vehicles necessitates comprehensive object detection.
• Existing vision-based systems struggle with detecting unknown object classes, posing a safety risk.

Purpose of the Study:

• To investigate the efficacy of anomaly detection using convolutional autoencoders for identifying unknown objects in agricultural fields.
• To compare the performance of various autoencoder models against a baseline object detector.

Main Methods:

• Convolutional autoencoders (AE, VQ-VAE, DAE, SSAE) were trained to reconstruct normal field patterns.
• Anomaly detection was performed by identifying objects with high reconstruction errors.
• Performance was evaluated against a YOLOv5 object detector using precision/recall AUC metrics.

Main Results:

• The semi-supervised autoencoder (SSAE) achieved a precision/recall AUC of 0.9353, outperforming other autoencoder models.
• SSAE's performance was comparable to the YOLOv5 object detector (PR AUC 0.9794).
• SSAE successfully detected unknown objects, a capability lacking in the YOLOv5 detector.

Conclusions:

• Anomaly detection using SSAE is a viable method for identifying unknown objects for autonomous agricultural vehicles.
• SSAE offers a robust solution for enhancing safety by detecting previously unseen obstacles.
• The approach addresses limitations of traditional object detectors in handling novel object classes.