IEEE Circuits and Systems Magazine - Q2 2021 - 52

based on the original and destination of the robot.
Broadly, prediction and routing modules are also included
here, where their outputs are fed into downstream
planning and control layers as input. The prediction
module is responsible for predicting the future
behavior of surrounding objects identified by the perception
layer. The routing module can be a lane-level
routing based on lane segmentation of the HD maps for
autonomous vehicles.
Planning and Control layers usually include behavioral
decision, motion planning and feedback control.
The mission of the behavioral decision module is to
make effective and safe decisions by leveraging all
various input data sources. Bayesian models are becoming
more and more popular and have been applied
in recent works [33], [34]. Among the Bayesian models,
Markov Decision Process (MDP) and Partially Observable
Markov Decision Process (POMDP) are the
widely applied methods in modeling robot behavior.
The task of motion planning is to generate a trajectory
and send it to the feedback control for execution. The
planned trajectory is usually specified and represented
as a sequence of planned trajectory points, and
each of these points contains attributes like location,
time, speed, etc. Low-dimensional motion planning
problems can be solved with grid-based algorithms
(such as Dijkstra [35] or A* [36]) or geometric algorithms.
High-dimensional motion planning problems
can be dealt with sampling-based algorithms, such as
Rapidly-exploring Random Tree (RRT) [37] and Probabilistic
Roadmap (PRM) [38], which can avoid the
problem of local minima. Reward-based algorithms,
such as the Markov decision process (MDP), can also
generate the optimal path by maximizing cumulative
future rewards. The goal of feedback control is
to track the difference between the actual pose and
the pose on the predefined trajectory by continuous
feedback. The most typical and widely used algorithm
in robot feedback control is PID.
While optimization-based approaches enjoy mainstream
appeal in solving motion planning and control
problems, learning-based approaches [39]-[43] are becoming
increasingly popular with recent developments
in artificial intelligence. Learning-based methods, such
as reinforcement learning, can naturally make full use of
historical data and iteratively interact with the environment
through actions to deal with complex scenarios.
Some model the behavioral level decisions via reinforcement
learning [41], [43], while other approaches directly
work on motion planning trajectory output or even
direct feedback control signals [40]. Q-learning [44],
Actor-Critic learning [45], policy gradient [38] are some
popular algorithms in reinforcement learning.
52
IEEE CIRCUITS AND SYSTEMS MAGAZINE
III. Perception on FPGA
A. Overview
Perception is related to many robotic applications where
sensory data and artificial intelligence techniques are
involved. Examples of such applications include stereo
matching, object detection, scene understanding, semantic
classification, etc. The recent developments in machine
learning, especially deep learning, have exposed
robotic perception systems to more tasks. In this section,
we will focus on the recent algorithms and FPGA implementations
in the stereo vision system, which is one of
the key components in the robotic perception stage.
Real-time and robust stereo vision systems are increasingly
popular and widely used in many perception
applications, e.g., robotics navigation, obstacle
avoidance [46] and scene reconstruction [47]-[49]. The
purpose of stereo vision systems is to obtain 3 D structure
information of the scene using stereoscopic ranging
techniques. The system usually has two cameras to
capture images from two points of view within the same
scenario. The disparities between the corresponding
pixels in two stereo images are searched using stereo
matching algorithms. Then the depth information can
be calculated from the inverse of this disparity.
Throughout the whole pipeline, stereo matching is
the bottleneck and time-consuming stage. The stereo
matching algorithms can be mainly classified into two
categories: local algorithms [50]-[56] and global algorithms
[57]-[61]. Local methods compute the disparities
by only processing and matching the pixels around
the points of interest within windows. They are fast
and computationally-cheap, and the lack of pixel dependencies
makes them suitable for parallel acceleration.
However, they may suffer in textureless areas and
occluded regions, which will result in incorrect disparities
estimation.
In contrast, global methods compute the disparities
by matching all other pixels and minimizing a global
cost function. They can achieve much higher accuracy
than local methods. However, they tend to come at high
computation cost and require much more resources due
to their large and irregular memory access as well as
the sequential nature of algorithms, thus not suitable for
real-time and low-power applications. Many research
works in stereo systems focus on the speed and accuracy
improvement of stereo matching algorithms, and
some of the implementations are summarized in Tab. I
B. Local Stereo Matching on FPGA
Local algorithms are usually based on correlation, where
the process involves finding matching pixels in the left
and right image patches by aggregating costs within a
SECOND QUARTER 2021
IEEE Circuits and Systems Magazine - Q2 2021

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q2 2021
