GStreamer Motion Detection Bin Performance

This section provides performance measurements for the rrmotiondetectionbin operating in various configurations. The following list describes different aspects of performance that were measured during the evaluation:

Frame rate: This indicates the number of frames that can be processed per second.
CPU utilization: This refers to the average percentage of CPU resources used by the pipeline.
GPU utilization: This refers to the average percentage of GPU resources used by the pipeline.

To measure these aspects RidgeRun's perf element and NVIDIA application tegrastats were used. The measurements were taken using Jetpack 4.5 on a Jetson Xavier NX and configured in power mode ID 5.The results of these measurements can be found in the following subsections.

Algorithms Configurations

The table below show the results of testing different motion detection algorithms, blob detection method and changing the processing format.

The motion detection algorithm was changed with the property motion_detector::motion and can take 2 values 'mog' or 'mog2'
The blob detection method was changed with the property cuda-blob-detection and can take 2 values true or false, true to use GPU based blob detection or false to use CPU based blob detection.
The processing format was changed with the property grayscale and can take 2 values: true or false, true to process the video in grayscale of false to process the value in RGBA.

The rest of the configurations uses the default values, noise reduction kernel size of 9 and sample frequency of 1.

4K Input Resolution

The measurements in this subsection were obtained from the following pipeline using a 4K input file:

gst-launch-1.0 filesrc location=/home/nvidia/street.mp4 ! qtdemux ! queue ! h264parse ! nvv4l2decoder ! queue ! rrmotiondetectionbin motion_detector::motion=$MOTION  cuda-blob-detection$BLOB_METHOD grayscale=$PROCESS_FORMAT ! perf ! fakesink sync=true

The rrmotiondetectionbin configuration were changed to determine its performance with different settings, changed with the corresponding element's properties.

Motion Detection Algorithm		Blob Detection Method		Processing Format		Framerate	CPU	GPU
mog	mog2	CPU	GPU	Grayscale	RGBA	Framerate	CPU	GPU
x		x		x		29.91	16.769	88.397
x			x	x		22.098	13.617	79.529
	x	x		x		29.921	15.963	75.115
	x		x	x		23.509	13.938	83.529
x		x			x	23.097	12	90.971

1080p Input Resolution

The measurements in this subsection were obtained from the following pipeline using a 4K input file downscaled to 1080p

gst-launch-1.0 filesrc location=/home/nvidia/street.mp4 ! qtdemux ! queue ! h264parse ! nvv4l2decoder !  nvvidconv ! "video/x-raw(memory:NVMM),width=1920,height=1080" !  queue ! rrmotiondetectionbin motion_detector::motion=$MOTION  cuda-blob-detection$BLOB_METHOD grayscale=$PROCESS_FORMAT ! perf print-cpu-load=true ! fakesink sync=true

The rrmotiondetectionbin configuration were changed to determine its performance with different settings, changed with the corresponding element's properties.

Motion Detection Algorithm		Blob Detection Method		Processing Format		Framerate	CPU	GPU
mog	mog2	CPU	GPU	Grayscale	RGBA	Framerate	CPU	GPU
x		x		x		29.927	10.423	48.391
x			x	x		29.921	8.74	51.807
	x	x		x		29.925	10	29.615
	x		x	x		29.921	10.423	40.629
x		x			x	29.928	10	51.08

Denoiser Kernel Size

To further analyze the impact of the denoiser we measured the performance for different kernel sizes, keeping the rest of configurations fixated as follows:

Motion detection algorithm: MOG
Blob detection: CPU
Sample Frequency: 1
Processing format: grayscale

4K Input Resolution

The following pipeline was used for these measurements:

gst-launch-1.0 filesrc location=/home/nvidia/street.mp4 ! qtdemux ! queue ! h264parse ! nvv4l2decoder ! queue ! rrmotiondetectionbin motion_detector::motion=mog motion_denoiser::size=$KERNEL_SIZE cuda-blob-detection=false ! perf print-cpu-load=true  ! fakesink sync=true

Denoiser Kernel Size	Framerate	CPU	GPU
5	29.906	18.076	86.577
9	29.91	16.769	88.397
11	29.924	16.153	84.462

1080p Input Resolution

The following pipeline was used for these measurements:

gst-launch-1.0 filesrc location=/home/nvidia/street.mp4 ! qtdemux ! queue ! h264parse ! nvv4l2decoder ! nvvidconv ! "video/x-raw(memory:NVMM),width=1920,height=1080" ! queue ! rrmotiondetectionbin cuda-blob-detection=false motion_detector::motion=mog motion_denoiser::size=$KERNEL_SIZE !  perf print-cpu-load=true ! fakesink sync=true

Denoiser Kernel Size	Framerate	CPU	GPU
5	29.927	9.777	40.16
9	29.927	10.423	48.391
11	29.925	10	37.923

Motion Sampling Frequency

This subsections shows the performance changes caused by different sampling frequencies, keeping the rest of configurations fixated as follows:

Motion detection algorithm: MOG
Blob detection: CPU
Denoise kernel size: 9
Processing format: grayscale

4K Input Resolution

The following pipeline was used for these measurements:

gst-launch-1.0 filesrc location=/home/nvidia/street.mp4 ! qtdemux ! queue ! h264parse ! nvv4l2decoder ! queue ! rrmotiondetectionbin motion_detector::motion=mog motion_detector::sample-frequency=$FREQ motion_denoiser::size=9 cuda-blob-detection=false ! perf print-cpu-load=true  ! fakesink sync=true

Sample Frequency	Framerate	CPU	GPU
1	29.91	16.769	88.397
2	30.232	11.44	56.292
4	30.448	8.72	40.739

1080p Input Resolution

The following pipeline was used for these measurements:

gst-launch-1.0 filesrc location=/home/nvidia/street.mp4 ! qtdemux ! queue ! h264parse ! nvv4l2decoder ! nvvidconv ! "video/x-raw(memory:NVMM),width=1920,height=1080" ! queue ! rrmotiondetectionbin cuda-blob-detection=false motion_detector::motion=mog motion_denoiser::size=9  motion_detector::sample-frequency=$FREQ !  perf print-cpu-load=true ! fakesink sync=true

Sample Frequency	Framerate	CPU	GPU
1	29.927	10.423	48.391
2	30.561	8.038	41.9
4	30.572	7.038	22.778

Previous: Performance

Index

Next: Troubleshooting

❯