Towards AI-Based Traffic Counting System with Edge Computing

Table 1

Experimental results about running models on Coral Dev Board.

Model	Model size (Mb)		Accuracy		FPS		Memory (%)	CPU usage (%)
	TFLite	Edge TPU TFLite	[email protected] : 0.95	[email protected]	TFLite	Edge TPU TFLite	TFLite and Edge TPU TFLite (%)	TFLite	Edge TPU TFLite
SSD MobileNet V1 300 × 300	5.37	5.75	0.445	0.768	2.6	27.5	17.8	26%	34.7%
SSD MobileDet 320 × 320	3.33	4.31	0.621	0.921	1.8	26.8	19.7	25.8%	33.3%
SSD MobileNet V2 320 × 320	4.58	5.18	0.462	0.794	3.2	25	18.3	26.2%	32.1%
SSD MobileNet V2 FPNLite 320 × 320	3.68	3.88	0.623	0.906	2.45	7.9	18	26.9%	26.7%
SSD MobileNet V2 FPNLite 640 × 640	4.26	4.74	0.754	0.951	0.63	2.3	19.1	26.4%	24.8%
SSD MobileNet V1 FPN 640 × 640	30	30.7	0.781	0.963	0.07	0.08	26.8	25.2%	25.1%