Real-Time Object Detection for the Running Train Based on the Improved YOLO V4 Neural Network

Table 4

Performance comparison of different network models.

Detection algorithm		Backbone	Params	mAP (%)	Training time (h)	FPS
Faster R–CNN		ResNet-50	25610269	90.02	31.72	1.2
SSD		VGG-16	24146894	84.20	4.26	22
YOLO V3		DarkNet-53	61952078	90.78	7.84	15.8
YOLO V4		CSPDarkNet-53	64429405	92.40	8.37	12.8
YOLO V4-tiny		CSPDarkNet-53-tiny	5924214	90.34	3.50	21.27
MYOLO-lite series	MV1-YOLO-lite	MobileNet V1	12754109	95.60	3.15	37.15
	MV2-YOLO-lite	MobileNet V2	10872893	94.17	2.87	36.71
	MV3-YOLO-lite	MobileNet V3	11791741	95.74	2.80	42.04