Table 2.

Summary and analysis of image fusion research methods.

Authors, Year	Key Research Points	Findings	Remarks
Lian et al. 2020 [15]	Multiple roadside camera perception data mapping to form a global semantic description.	The detection time was increased by about 45%, and the detection accuracy was increased by about 10%.	Distributed interactive fusion deployment of sensors for a wider range of cooperative perception without increasing the time cost of computing.
Löhdefifink et al. 2020 [21]	Used a lossy learning method for image compression to relieve the pressure on wireless communication channels.	Image compression requirements were high, and late fusion results of segmentation mask cascades were optimal.	The transmission of processed data can effectively reduce the load on the wireless channel.
Lv et al. 2021 [23]	Based on the separation principle of static background and dynamic foreground, the dynamic foreground was extracted, and the static background and dynamic foreground were re-fused by a generative adversarial network.	The processing time of perceptual information was reduced to 27.7% of the original.
Xiao et al. 2018 [24]	A bird’s-eye view generated by integrating the perception information of other vehicles expanded the perception range, shared the processed image information, and reduced the network burden.	Solved the problem of obstacle occlusion and reduced the transmission of data volume.	Perception range will be affected by communication distance.
Sridhar1 et al. 2018 [25]	Utilized image feature point matching for data fusion to form vehicle cooperative perception with a common field of view.	Fusion of perception information from other vehicles and conversion to its own coordinate system.	Cooperative perception can effectively expand the perception range of vehicles.
Liu et al. 2018 [27]	Used feature point matching to estimate geometric transformation parameters to solve perception blind spots in congestion.	The intersection over union value was increased by 2~3 times.	Effectively solved the obstacle occlusion, but ignored the problem of viewing angle.