Page 69 - Read Online

Zhuang et al. Intell Robot 2024;4(3):276-92 I http://dx.doi.org/10.20517/ir.2024.18 Page 290

throughput data for the model. Thus, underutilized GPU resources were solved, and the inference of the net-
work framework was optimized to improve the inference efficiency. We build a visual system interface using
the PyQT5 library [Figure 9].

Figure 10 demonstrates the workflow of the whole system. Firstly, the model is trained on the computer. Af-
terward, the trained model is solidified and deployed on Jetson Nano. The camera converts the acquired road
conditions into images and inputs them into Jetson Nano. After the vehicle and pedestrian detection, Jetson
Nano gives instructions such as braking and turning to the car.

6.CONCLUSION
This paper introduces an in-vehicle IR target detection method built on an improved YOLOv4 model, which
integrates the IE-CGAN inversion algorithm to pre-process IR images. This integration enhances both image
quality and detection performance. An IE-CGAN inversion algorithm is used instead of conventional meth-
ods to pre-process the IR images. Additionally, considering that the algorithm requires being deployed on the
edge device, this study concentrates on improving the system’s processing speed for IR images, so the back-
bone network of the YOLOv4 model is replaced from CSPDarknet53 to MobileNetV3, improving processing
speed and efficiency. However, the dataset we used has limited diversity in image types and needs more gen-
eralizability in background models. Moreover, the model’s ability to generalize requires further improvement.
Our future work will focus on expanding the dataset to include a broader range of IR images, enhancing the
system’s robustness and generalizability across different scenarios. Moving forward, we remain committed to
addressing the current limitations and enhancing the system’s performance and generalizability in future work.

DECLARATIONS
Authors’ contributions
Made substantial contributions to the conception and design of the study and performed data analysis and
interpretation: Zhuang T, Liang X
Performed data acquisition and provided administrative, technical, and material support: Xue B, Tang X

Availability of data and materials
The FILR datasets used in this article can be found at https://www.flir.com/oem/adas/adas-dataset-form/.

Financial support and sponsorship
This work was supported by the National Natural Science Foundation of China (Grant No.62001173), the
Project of Special Funds for the Cultivation of Guangdong College Students’ Scientific and Technological In-
novation (“Climbing Program” Special Funds) (Grant Nos. pdjh2022a0131 and pdjh2023b0141).

Conflicts of interest
All authors declared that there are no conflicts of interest.

Ethical approval and consent to participate
Not applicable.

Consent for publication
Not applicable.

Copyright
© The Author(s) 2024.