Table 1.

Comparison of proposed method and previous study results.

Category	Method		Advantages	Disadvantages
Multiple camera-based method	Using visible light and FIR cameras [27]	Spatial-temporal filtering, seeded region growing, and min-max score fusion	Uses data from two cameras to improve human detection accuracy	-Correspondence points must be manually set between two cameras for calibration -Requires a sequence of image frames -Difficult to use in most normal surveillance environments because of high cost of FIR cameras -Images from two cameras must be processed, which takes a long time, and the processing speed is lowered if many objects are detected
Single camera-based methods	Using IR camera (NIR or FIR camera)	GMM [14], SVM classifier with feature vector from human region [15] and by HOG [16]	Uses one camera, which eliminates the need for calibration, and has a faster processing time than multiple camera-based methods	-Can only be used in a fixed camera environment [14] -If an NIR camera is used, an additional NIR illuminator must be used. NIR illuminators are limited in terms of their illumination angle and distance, and the illuminator power must be adaptively adjusted for near and far objects [16] -FIR cameras are expensive, and their image resolution is much lower than visible light cameras. Thus, there are few features that can be captured in a human area during human detection at a long distance [14,15,16]
	Using visible light camera	Uses local change in contrast over time [28,29]	Uses low-cost visible light cameras	-Can only be used in a fixed camera environment -Has trouble detecting humans who are standing still -Must use continuous video frames, which requires a fast capture speed, and it processes multiple images, which increases the processing time
		Histogram processing or intensity mapping-based image enhancement [32,33,34,35,36,37]	Uses low-cost visible light cameras	-Has only produced experimental results for raising the visibility through image enhancement, with no results produced for human detection in nighttime images
		Denoising and image enhancement [38,39]	Effectively removes noise that occurs during image enhancement	-Because denoising requires many operations, the processing time is long compared to histogram methods -Has only produced experimental results for raising visibility through image enhancement, with no results produced for human detection in nighttime images
		CNN (proposed method)	Independently processes single images. Thus, even stationary objects can be detected. Can be used with moving or fixed cameras	-Adequate data and time are required to train CNN