Page 9 - Read Online

Lei et al. Intell Robot 2022;2(4):313­32 I http://dx.doi.org/10.20517/ir.2022.18 Page 315

algorithm [21] . However, the design and implementation of path planning in a broiler barn involve multiple
aspects for mortality removal. Especially with the large scale of modern broiler barns (e.g. 12 m × 200 m),
poultry mortality is widely distributed anywhere in the barn, which increases the difficulty of detecting and
removing dead broilers. Therefore, we propose a system for searching and removing broiler mortality with
two robots. One is a broiler mortality detection robot and the other is a broiler mortality removal robot. The
major function of the detection robot is to precisely and efficiently search the whole broiler barn and indi-
cate the position of dead broilers. Thus, coverage path planning (CPP) algorithms need to be introduced to
assist robots in search and exploration tasks, such as bio-inspired neural networks [22] , Boustrophedon grid
decomposition [23] , deep learning [24,25] , etc.

Luo and Yang [22] developed the bio-inspired neural network (BNN) method to navigate robots to perform
complete coverage path planning (CCPP) while avoiding obstacles within dynamic environments. The robot
is attracted to unscanned areas and repelled by the previous scanned areas or obstacles based on the neu-
ral activity through the BNN model. The next position of the robot depends on the current position of the
robot and neuron activity associated with its current position. Zhu et al. proposed a complete coverage path
planning model using Glasius BNN, which is extended to multiple robots to lower the overall search time
and improve efficiency [19] . Unlike the BNN approach, the boundary representation model that defines the
workspace is adopted by the Boustrophedon cellular decomposition (BCD) method and deep reinforcement
learning approach (DRL). The BCD method proposed by Acar and Choset [23] decomposes the environment
into many line scan partitions explored through a back-and-forth path (BFP) in the same direction. BCD is an
effective CCPP method for more diverse, non-polygonal obstacles in a workspace. In trapezoidal decomposi-
tion as a cell, it is covered in back-and-forth patterns. For a complex configuration space with irregular-shaped
obstacles, BCDneedstoconstructagraphthatrepresentstheadjacency connectionsofthecellsin theBoustro-
phedon decomposition. Similarly, Nasirian et al. utilized traditional graph theory to segment the workspace
and proposed a deep reinforcement learning approach to solve the CCPP problem in a complex workspace [24] .
Lei et al. proposed a deep learning method to detect the workspace and generate turning waypoints for the
robot to complete the coverage of the entire workspace [25] . The trajectory generated by the above-presented
algorithms completely covers the workspace in light of the size of the robot, which is more suitable for a large-
sized workspace with a broad sensing range. However, with the limited sensing range of robots in broiler barns,
a complete coverage trajectory requires a large amount of time and energy, which may not be affordable to the
farmers.

The methods of object detection have been extensively studied, and many object detectors are utilized in many
agricultural settings. Li et al. proposed a fast region-based CNN (R-CNN) detector with high accuracy to
adapt to the illumination of different colored lights in the farm and detect the cage-free floor eggs [26] . A mask
R-CNN object detector was developed by Li et al. to detect hen preening behaviors in barns to automatically
monitorpoultrybehaviors,judgethecomfortlevelofhens,andassistwelfare-orientedpoultrymanagement [27] .
Bochkovskiy et al. proposed a YOLO V4 object detector, which is a deep learning-based object detection
technique with high accuracy [28] . It is an effective object detector to localize objects in real-time processing.

1.3. Proposed methods and original contribution
A directed coverage path planning (DCPP) fused with an informative planning protocol (IPP) is proposed to
efficiently search the entire barn. The broiler barn is first decomposed into large grids based on the workspace
dimensions to obtain overall global trajectories with coverage directions. IPP continues to rapidly achieve spa-
tial coverage with the least estimation uncertainty in the decomposed grids. Flexible and efficient trajectories
are formed by utilizing historical information of broiler mortality spatial distribution [29] and the direction in-
formation from the previous steps. With the assistance of current state-of-the-art computer vision algorithms
in precision agriculture, such as deep learning techniques [30] , the robot could detect dead birds in the grids
of vision range and find the mortality location for the removal robot. Such a comprehensive path planning