This Python sample may be slow depending on the event rate of the scene and the configuration of the algorithm. We provide it to allow quick prototyping. For better performance, look at the corresponding C++ sample.

Object Counting using Python

The Analytics API provides algorithms to count small fast moving objects.

The sample shows how to use the python bindings of Metavision Analytics SDK to count and display the objects passing in front of the camera.

Objects are counted per lines (by default, 4 horizontal lines are used), and we expect objects to move from top to bottom, as in free-fall. Each object is counted when it crosses one of the horizontal lines. The number of lines and their positions can be specified using command-line arguments.

The source code of this sample can be found in <install-prefix>/share/metavision/sdk/analytics/python_samples/metavision_counting when installing Metavision SDK from installer or packages. For other deployment methods, check the page Path of Samples.

Expected Output

Metavision Counting sample visualizes the events (from moving objects), the lines on which objects are counted, and the total object counter:

Setup & requirements

To accurately count objects, it is very important to fulfill some conditions:

  • the camera should be static and the object in focus

  • there should be good contrast between the background and the objects (using a uniform backlight helps to get good results)

  • set the camera to have minimal background noise (for example, remove flickering lights)

  • the events triggered by an object passing in front of the camera should be clustered as much as possible (i.e. no holes in the objects to avoid multiple detections)

Also, we recommend to find the right objective/optics and the right distance to objects, so that an object size seen by the camera is at least 5 pixels. This, together with your chosen optics, will define the minimum size of the objects you can count.

Finally, depending on the speed of your objects (especially for high-speed objects), you might have to tune the sensor biases to get better data (make the sensor faster and/or less or more sensitive).

How to start

To start the sample based on the live stream from your camera, run:





To start the sample based on recorded data, provide the full path to a RAW file (here, we use a file from our Sample Recordings):


python3 -i 80_balls.raw


python -i 80_balls.raw

To check for additional options:


python3 -h


python -h

Code Overview


Metavision Counting sample implements the following pipeline:


Optional Pre-Processing Filters/Algorithms

To improve the quality of initial data, some pre-processing filters can be applied upstream of the algorithm:


These filters are optional: experiment with your setup to get the best results.

Counting Algorithm

This is the main algorithm in this sample. The algorithm is configured to count objects of a given size passing from top to bottom in front of the camera. To create an instance of metavision_sdk_analytics.CountingAlgorithm, we first need to gather some configuration information, such as the size of the objects to count, their speed and their distance from the camera. The size of those objects in the camera’s image plane depends on the optic used, their distance to the camera and their speed. The metavision_sdk_analytics.CountingCalibration class allows to compute these parameters and pass them to the algorithm. Once we have a valid calibration, we can create an instance of metavision_sdk_analytics.CountingAlgorithm:

# Counting Calibration (Get optimal algorithm parameters)
cluster_ths, accumulation_time_us = CountingCalibration.calibrate(
    width=width, height=height, object_min_size=args.object_min_size,
    object_average_speed=args.object_average_speed, distance_object_camera=args.distance_object_camera)

The metavision_sdk_analytics.CountingAlgorithm relies on the use of lines of interest to count the objects passing in front of the camera and produces 4 information through its output callback: (ts, global_counter, last_count_ts, local_counters). Local counters (i.e. one per line) are incremented every time an object crosses their corresponding line, while the global counter is the maximum of all the local counters. These counters are not reset between two calls but updated throughout the sequence. The algorithm is implemented in an asynchronous way which allows to retrieve new counters estimations at a fixed refresh rate rather than getting them for each processed buffer of events. As these counters are mainly used for visualization purposes, the asynchronous approach has been proven to be more efficient in this case.

Frame Generation

At this step, we generate an image that will be displayed when the sample is running. In this frame are displayed:

  • the events

  • the lines of interest used by the algorithm

  • the global counter

The metavision_sdk_core.OnDemandFrameGenerationAlgorithm class allows to buffer input events (i.e. metavision_sdk_core.OnDemandFrameGenerationAlgorithm.process_events()) and generate an image on demand (i.e. metavision_sdk_core.OnDemandFrameGenerationAlgorithm.generate()). Once the event image has been generated, the counting related overlays (i.e. lines and counter) are rendered using the metavision_sdk_analytics.CountingDrawingHelper class.

As the output images are generated at the same frequency as the counts produced by the metavision_sdk_analytics.CountingAlgorithm, the image generation is done in the metavision_sdk_analytics.CountingAlgorithm’s output callback.

The event processing is done while iterating over metavision_core.event_io.EventsIterator:

# Process events
for evs in mv_iterator:
    # Dispatch system events to the window

    # Process events
    if filtering_algorithms:
        filtering_algorithms[0].process_events(evs, events_buf)
        for filter in filtering_algorithms[1:]:

    if counting_gui.should_close():


Finally, the generated frame is displayed on the screen. The following image shows an example of output:

Expected Output from Metavision Counting Sample