TractorEYE: Vision-based Real-time Detection for Autonomous Vehicles in Agriculture


Peter Hviid Christiansen


Autonomous Tractor, Deep Learning, Machine Learning, Obstacle detection, Object detection, Computer Vision, Sensor Fusion, FieldSAFE, Agriculture, Mapping, Sensor Registration, Sensor Calibration


Agricultural vehicles such as tractors and harvesters have for decades been able to navigate automatically and more efficiently using commercially available products such as auto-steering and tractor-guidance systems. However, a human operator is still required inside the vehicle to ensure the safety of vehicle and especially surroundings such as humans and animals. To get fully autonomous vehicles certified for farming, computer vision algorithms and sensor technologies must detect obstacles with equivalent or better than human-level performance. Furthermore, detections must run in real-time to allow vehicles to actuate and avoid collision.

This thesis proposes a detection system (TractorEYE), a dataset (FieldSAFE), and procedures to fuse information from multiple sensor technologies to improve detection of obstacles and to generate a map. 
TractorEYE is a multi-sensor detection system for autonomous vehicles in agriculture. The multi-sensor system consists of three hardware synchronized and registered sensors (stereo camera, thermal camera and multi-beam lidar) mounted on/in a ruggedized and water-resistant casing. Algorithms have been developed to run a total of six detection algorithms (four for rgb camera, one for thermal camera and one for a Multi-beam lidar) and fuse detection information in a common format using either 3D positions or Inverse Sensor Models. A GPU powered computational platform is able to run detection algorithms online. For the rgb camera, a deep learning algorithm is proposed DeepAnomaly to perform real-time anomaly detection of distant, heavy occluded and unknown obstacles in agriculture. DeepAnomaly is -- compared to a state-of-the-art object detector Faster R-CNN -- for an agricultural use-case able to detect humans better and at longer ranges (45-90m) using a smaller memory footprint and 7.3-times faster processing. Low memory footprint and fast processing makes DeepAnomaly suitable for real-time applications running on an embedded GPU. 
FieldSAFE is a multi-modal dataset for detection of static and moving obstacles in agriculture. The dataset includes synchronized recordings from a rgb camera, stereo camera, thermal camera, 360-degree camera, lidar and radar. Precise localization and pose is provided using IMU and GPS. Ground truth of static and moving obstacles (humans, mannequin dolls, barrels, buildings, vehicles, and vegetation) are available as an annotated orthophoto and GPS coordinates for moving obstacles. 

Detection information from multiple detection algorithms and sensors are fused into a map using Inverse Sensor Models and occupancy grid maps. 
This thesis presented many scientific contribution and state-of-the-art within perception for autonomous tractors; this includes a dataset, sensor platform, detection algorithms and procedures to perform multi-sensor fusion. Furthermore, important engineering contributions to autonomous farming vehicles are presented such as easily applicable, open-source software packages and algorithms that have been demonstrated in an end-to-end real-time detection system. The contributions of this thesis have demonstrated, addressed and solved critical issues to utilize camera-based perception systems that are essential to make autonomous vehicles in agriculture a reality.



7 november 2018

Detaljer om denne monografi

ISBN-13 (15)