Test results

Test machine specification

CPU	RAM	Graphics Card
AMD Ryzen 9 5950X @ 3.4 GHz: 16 cores (32 threads)	118GB DDR4	NVIDIA GeForce GTX 1080 Ti

Testing conditions

For testing, we took a video of a corridor with a flow of people moving at a speed of 5-10 people/sec. The volume of the database for identification is 249 persons.

Metrics

Metric name	Description
Maximum	Maximum value of parameters (CPU load, GPU load, amount of RAM, amount of VRAM) for the entire test time
95th percentile	CPU is loaded at no more than the specified value for 95% of the test time GPU is loaded at no more than the specified value for 95% of the test time The amount of RAM used does not exceed the specified value for 95% of the test time The amount of VRAM used does not exceed the specified value for 95% of the test time
Median	CPU is loaded at no more than the specified value for 50% of the test time GPU is loaded at no more than the specified value for 50% of the test time The amount of RAM used does not exceed the specified value for 50% of the test time The amount of VRAM used does not exceed the specified value for 50% of the test time
Recall	Percentage of identified faces / actions
Precision	Percentage of accurate face identifications / accurate classifications of actions
Max latency	Max time spent between event generation and submission
Total events out of 117	This number shows how many identification events were generated during the test (Maximally the test result should contain 117 identification events)

Face recognition testing

Test results

	Number of video streams	1	5	8	10	12	15
Maximum	Number of CPU cores*	1	6	10	13	15	19
	GPU load	91%	95%	96%	95%	97%	95%
	Amount of RAM	3.83 GB	6.80 GB	9.27 GB	10.76 GB	12.24 GB	14.09 GB
	Amount of VRAM	2.15 GB	2.41 GB	2.86 GB	2.86 GB	2.86 GB	2.87 GB
95th percentile	Number of CPU cores*	1	6	10	13	15	19
	GPU load	17%	65%	71%	71%	68%	59%
	Amount of RAM	2.97 GB	6.68 GB	9.15 GB	10.26 GB	11.50 GB	13.35 GB
	Amount of VRAM	2.09 GB	2.41 GB	2.86 GB	2.86 GB	2.86 GB	2.87 GB
	Recall	99.10%	99.10%	99.10%	98.30%	98.30%	95.70%
	Precision	100%	100%	100%	100%	100%	100%
	Max latency (sec)	12	13	13	14	15	19
	Total events out of 117	116	116.2	116.25	114.9	114.6	112.3

*/ The number of cores refers to the number of logical cores with Hyper-Threading (HT) or Simultaneous Multithreading (SMT).

Conclusion

For the most high-load use case, "Safe City," resource allocation is as follows:

• With GPU: Approximately 1.4 CPU cores per video stream are required.
• Without GPU: The load increases to 2.2 CPU cores per video stream.

The number of people in the frame does not affect resource consumption.

Human Action Recognition (HAR) Testing on GPU

The quality tests for Human Action Recognition (HAR) were performed with an average of 4 people in the frame.

Conclusion

Resource consumption per person in the frame: 0.25 CPU, 0.9 GB RAM, 0.7 GB VRAM, 2.5 TFLOP.

Algorithm accuracy:

	Fight	Sit	Lie
Recall	74%	80%	64%
Precision	98%	99%	90%

Skeleton tracking testing

Test results

	Number of video streams	1	2	3	5	8
Maximum	Number of CPU cores*	3	3	5	7	10
	GPU load	48%	84%	92%	90%	98%
	Amount of RAM	1.48 GB	1.98 GB	2.60 GB	3.71 GB	5.44 GB
	Amount of VRAM	0.55 GB	0.87 GB	1.23 GB	1.95 GB	3.34 GB
95th percentile	Number of CPU cores*	2	3	5	7	10
	GPU load	29%	47%	74%	85%	95%
	Amount of RAM	1.48 GB	1.98 GB	2.47 GB	3.65 GB	4.82 GB
	Amount of VRAM	0.52 GB	0.81 GB	1.23 GB	1.95 GB	3.34 GB
	Recall	87.18%	87.18%	87.18%	78.97%	65.38%
	Precision	82.93%	86.20%	83.62%	89.01%	91.11%

*/ The number of cores refers to the number of logical cores with Hyper-Threading (HT) or Simultaneous Multithreading (SMT).

Conclusion

Resource consumption per video stream: 1.5 CPU cores, 1 GB RAM, 0.5 GB VRAM, 2.6 TFLOP (Independent of the number of people in the frame).