Page 56 - Read Online

Page 277 Zhuang et al. Intell Robot 2024;4(3):276-92 I http://dx.doi.org/10.20517/ir.2024.18

1. INTRODUCTION
With the speedy development of transportation systems, the increasing number of vehicles is leading to more
[1]
traffic problems, and the risk of traffic accidents continues to rise . According to research, around 1.35 mil-
[2]
lion people die globally because of traffic accidents each year . Financial costs due to traffic accidents add
[3]
up to 1%-3% of the world’s gross domestic product . A large proportion of traffic accidents occurs under
[4]
reduced visibility conditions when the view of drivers is greatly limited . How to minimize the occurrence
of traffic accidents has been a hot issue, and advanced driver assistance systems (ADAS) are considered a fea-
[5]
sible way to achieve this goal . As an important way to obtain information from the external environment,
machine vision is considered to be the core technology of ADAS. In recent years, infrared (IR) imaging tech-
nology has been adopted to obtain information, which makes it possible to implement in-vehicle IR target
[6]
detection systems . IR images are not easily affected by changes in external light, which is an outstanding
advantage compared with visible images. These images can provide more valuable information for intelligent
[7]
automotive systems , especially in low-visibility weather such as heavy rain and fog. Therefore, the applica-
tion of IR imaging technology is a considerable choice to make the safe driving system more reliable. With the
development of technology, the research of computer vision is iterated year by year. In 2005, Dalal and Triggs
[8]
proposed the Histogram of Gradient (HOG) detector which became an important improvement in Scale In-
variant Feature Transform and Shape Contexts at that time. Related technologies are widely used in computer
vision applications and lay an important foundation for many later detection methods. In 2014, Girshick et al.
[9]
proposed the Region with CNN features (R-CNN) , which selects possible object boxes from a set of object
candidate boxes through the selective search algorithm and then resizes the images in these selected object
boxes to a fixed-size image. Later the algorithm feeds them to the trained CNN model to extract features and
finally sends the extracted features to the classifier to predict whether the image in the object box has a target
to be detected. And further, predict which category the detection target belongs to. In 2016, Redmon et al.
proposed you only look once (YOLO) v1 [10] , which is the first stage of the deep learning detection algorithm.
Its detection speed is very fast; the idea of the algorithm is to divide the image into multiple grids, and then
predict the bounding box for each grid at the same time and give the corresponding probability. Based on this
idea, YOLOv1 has been continuously developed into v2, v3, v4, v5 and other versions. In 2018, Law and Deng
proposed the CornerNet algorithm [11] . As the pioneer of the Anchor technology route, the network uses a
new target detection method, which transforms the detection of the target bounding box by the network into
a pair of key points (the upper left corner and the lower right corner).

This work proposes a method to develop an in-vehicle IR target detection system. The key contributions of
this paper are:

• Building on the Image Enhancement Conditional Generative Adversarial Network (IE-CGAN) proposed by
Kuang et al., we introduced an innovative improvement, resulting in the IE-CGAN inversion algorithm [12] .
This algorithm enhances input images, replacing traditional pre-processing methods.

• The YOLOv4 model is optimized by replacing its backbone network, CSPDarknet53, with MobileNetV3.
This replacement has been shown to effectively enhance the system’s real-time detection capabilities while
maintaining high detection accuracy.

• The model is deployed on a Jetson Nano, an edge device with limited hardware resources, culminating in a
fully integrated system that combines both hardware and software.

2. RELATED WORKS
In ADAS, detecting vehicles and pedestrians is a core task. Currently, many effective methods have been pro-
posed, including two aspects: the YOLO series algorithm and the IR target recognition part. Among them,