AI On the Edge, Python

AI on the Edge LESSON 28: Use Pan Tilt Camera to Track Object of Interest in OpenCV

Hey everyone, Paul McWhorter here from TopTechBoy.com. Welcome back to our channel, where we don’t just write abstract software—we build real, physical, intelligent machines. Go ahead and grab yourself a nice hot cup of coffee or a big glass of iced tea, because today we are closing the loop between the digital world of computer vision and the physical world of robotics.

In our last lesson, we successfully taught OpenCV how to find a specific color, isolate the largest shape, and draw a beautiful green bounding box around it. That was great, but it had a massive limitation: if your object moved off the edge of the frame, it was gone forever. The camera just sat there, blind and helpless.

Today, we change that. We are taking that tracking data from our code and using it to command a physical pan-tilt mechanism powered by two servos on our Fusion Hat. By the time we finish today, your camera will physically turn, tilt, and hunt down your object, keeping it locked dead center in the middle of your video feed.

The Big Leap: Closing the Loop

What we are building today is a foundational concept in automation engineering known as a feedback control loop.

Up until now, your camera was an open-loop observer. It saw things, but it couldn’t react physically. To make an autonomous tracking system, we need to implement a simple pipeline:

  1. Sense: The camera captures the image frame.

  2. Think: OpenCV finds the target object and calculates its position.

  3. Act: The script commands the hardware servos to move the camera mount to correct any positioning errors.

The Mathematics of the Target Error

To make a camera track an object, we have to define what “perfect tracking” looks like to a computer. Perfect tracking means the center of our tracked object is sitting exactly at the center of our video frame.

Because we are running our camera at a crisp resolution of 1280×720, the mathematical center of our universe is fixed. We calculate our frame’s horizontal and vertical centers by dividing our dimensions in half. This gives us a permanent anchor point right in the middle of our grid.

When an object appears on screen, our contour detection gives us its bounding box. We calculate the exact center of that box by taking its starting coordinate and adding half of its width and height. Now we have two sets of coordinates:

  • Where we want the object to be (The Frame Center).

  • Where the object actually is (The Box Center).

The difference between where the object is and where it belongs is called the Error Signal. We calculate an X Error and a Y Error by simply subtracting the frame center from the box center.

Managing Jitter with a Control Deadband

If our object is perfectly centered, our Error is zero. If the object moves to the right, the X Error becomes a positive number. If it moves to the left, it becomes a negative number. The same logic applies vertically to our Y Error.

Now, you might think we should tell the servos to move every single time the Error is anything other than zero. But remember what we learned about camera sensors: pixels dance, light fluctuates, and your calculations will always have a tiny amount of natural mathematical noise. If you try to correct for every single fractional pixel change, your servos will constantly buzz, twitch, and jitter themselves to death.

To fix this, we implement an engineering safety margin called a Deadband. In this lesson, we establish a 40-pixel safety zone around the center of the frame.

  • If the object is within 40 pixels of the center, the error is too small to care about, and we tell the servos to sit perfectly still.

  • The moment the object drifts outside that 40-pixel window, our control logic triggers.

If the X Error is greater than 40, we decrement our pan angle by one degree to turn the camera toward the target, pass that new angle to our servo handler, and pause for a tiny fraction of a second (20 milliseconds) to give the mechanical gears time to physically move. If it’s negative, we increment the angle. We apply the exact same behavioral logic to our tilt servo using the Y Error.

Visualizing the System Matrix

To help us calibrate and troubleshoot this system, we overlay clear visual indicators directly onto our live video feed:

  • The Reticle: We draw a solid blue dot directly at our fixed frame center. This acts as our tracking crosshair.

  • The Target: We draw a large red circle directly over the center of our moving object’s bounding box.

When your system is working properly, you can physically watch the machine think. As you move an object around, the red circle moves away from the blue dot, the error threshold trips, the servos kick in, and the camera moves until the red circle swallows the blue dot once again.

Your Homework Assignment

You guys know the drill: watching me build a tracking rig doesn’t make you an automation engineer. You have to write the logic, feel the hardware move, and tune it yourself.

Here is your homework challenge for Lesson 28: Right now, our tracking logic uses what is called an incremental step controller. No matter how far away the object is from the center, the camera always moves at the exact same speed—one lazy degree at a time. If you move your target slowly, the camera keeps up. If you snap your target quickly across the room, the camera falls behind and loses it because it can’t accelerate.

Your assignment is to upgrade this control loop. Instead of stepping by a hardcoded value of 1, I want you to make the servo adjustment step proportional to the size of the error. If the object is close to the center, it should move gently by a fraction of a degree. If the object takes off like a rocket and creates a massive error signal, the camera should aggressively throw the servos open to catch up instantly.

Get your proportional tracking loops tuned, shoot a video showing your camera tracking a fast-moving object smoothly, upload it to YouTube, and share your link down in the comments below. See you guys in the next lesson!

AI On the Edge, Raspberry Pi

AI on the Edge LESSON 27: Track Objects of Interest in OpenCV Using Contours

AI on the Edge LESSON 27: Track Objects of Interest in OpenCV Using Contours

Hey everyone, Paul McWhorter here from TopTechBoy.com. Welcome back to our channel, where we learn to build real, intelligent systems on edge hardware. Grab yourself a nice hot cup of coffee or a cold glass of iced tea, because today we are taking a massive leap forward in our computer vision journey.

Up until now, we have learned how to configure our cameras, calculate frame rates smoothly, and isolate specific objects based on color using the HSV color space. We built beautiful masks and composite images that show only our target color. But let’s be honest with ourselves: a mask is just a collection of white pixels on a black screen. The computer doesn’t actually know where the object is, how big it is, or how to follow it if it moves.

In this lesson, we are going to fix that. We are going to teach the machine to look at our mask, isolate the single biggest shape of interest, ignore the background noise, and draw a real-time bounding tracking box around it. This is true object tracking.

The Core Concept: What is a Contour?

Think of a contour as a mathematical boundary line. When OpenCV looks at a binary mask (where your target object is white and everything else is black), a contour is the continuous line that traces the outer edge of that white shape.

The beauty of contours is that they turn a chaotic cloud of thousands of isolated pixels into structured, manageable vector shapes. Once OpenCV finds these shapes, it can calculate their physical properties, such as their area, perimeter, and exact center.

The Three Steps to Algorithmic Object Tracking

To turn a raw camera frame into a fully tracked target, our script follows a strict three-part engineering pipeline inside our main execution loop:

1. Extracting Every Boundary

First, we pass our binary mask into OpenCV’s contour detection engine. We configure it to use external retrieval, meaning it will ignore any hollow holes inside the object and only trace the outermost boundary. It returns a list of every single contour it finds in the frame.

2. Hunting for the Largest Target

In the real world, your camera view is never perfectly clean. Even with an excellent HSV color mask, you will get random speckles, reflections, or background noise showing up as tiny white dots on your mask. If we tried to track everything, our program would lose its mind. To solve this, we use a Python maximization function to scan our list of contours and extract the absolute largest one based on its physical area.

3. Setting an Area Noise Floor

Even after finding the largest contour, what happens if your object completely leaves the camera view? The largest remaining “object” might be a tiny, single-pixel spec of static noise on the edge of the screen. To prevent our tracking box from jumping around erratically, we establish a strict structural threshold—a noise floor. If the area of the largest contour isn’t big enough to confidently be our target, we ignore it completely.

Drawing the Bounding Box

Once we have successfully isolated our valid, large contour, we don’t just want to draw a messy, squiggly line around it. We want clean coordinates that an automation system or a robotic pan-tilt kit could actually use to follow the target.

We pass our largest contour into a bounding rectangle function. OpenCV automatically calculates the exact mathematical limits of that shape and returns four precise numbers:

    • X: The horizontal starting pixel coordinate of the object.

    • Y: The vertical starting pixel coordinate of the object.

    • W: The total width of the object in pixels.

    • H: The total height of the object in pixels.

With those four dimensions locked down, we use a standard drawing function to overlay a crisp, green rectangle directly onto our live color camera feed. Now, as you move your object around the room, the box follows it dynamically, tracking its position in real time at high frame rates.

Note you will have to tune the LC and UC parameters for your object of interest, as we showed last week.

 

AI On the Edge, Raspberry Pi

AI on the Edge LESSON 26: Understanding the HSV Color Space in OpenCV

Hey guys, welcome back to the channel. If you’ve been following along, you know we’ve been pushing our hardware absolutely down into the dirt. We’ve been running large language models right on the edge, pushing our boards hot and heavy until the silicon is screaming and the thermal throttling flags are popping up all over the place.

But today, we are stepping away from the heavy-compute server terminals, and we are getting back to our roots: Real-Time Computer Vision and Embedded Control. In our previous lessons, we learned how to hook up our high-speed camera, capture raw frames, and interact with individual pixels using standard RGB/BGR math. But today, we are going to look under the hood of a completely different way of representing color: The HSV Color Space (Hue, Saturation, Value).

If you try to track objects or isolate specific colors in the traditional RGB world, you are going to pull your hair out. The moment a shadow hits your object or the room lighting changes, your Red, Green, and Blue values completely collapse. By shifting our mathematics into the HSV space, we can lock onto a color’s pure identity regardless of whether it is sitting under a bright laboratory spotlight or a dim shadow.

Not only are we going to capture and process these video streams at a smooth-as-silk 60 frames per second, but we are also going to translate that raw visual math directly into the physical world. We are using our trusty SunFounder Fusion HAT+ to dynamically pulse an external RGB LED, matching its brightness and color hue perfectly to whatever pixel your mouse is clicking on in real-time.

Let’s look at the blueprint to make this happen.

The Complete Python Code

Here is the clean, un-guardrailed Python script for today’s lesson. Paste this directly into your local terminal workspace. No bloated libraries, no unnecessary frameworks—just pure, deliberate engineering.

Under the Hood: How the Code Works

1. The Real-Time Telemetry Smooth Filter

Look closely at how we calculate our frames-per-second metric inside the main processing loop:

If you simply print out the raw math of 1 / deltaT, your numbers on the screen are going to jump all over the place like a wild animal. By applying a 95% historical weight and a 5% instant weight, we create a low-pass software filter that smoothly tracks our true hardware operational speed without erratic layout jitter.

2. The Mouse Vector and BGR Array Sequence

When your mouse triggers an event over the window, OpenCV passes us the standard coordinate pairs (x, y). But remember: inside a NumPy data structure, images are structured as Rows first, then Columns. That means when you slice into your image array to read a pixel’s color values, you must pass the parameters as frame[y, x]. If you pass it as [x, y], your program is going to index out of bounds and crash hard.

Furthermore, always remember that OpenCV handles colors in a BGR (Blue, Green, Red) sequence, not RGB. When we extract those elements, they unpack straight into valB, valG, valR.

3. Masking and Bitwise Isolation

To lock onto our target color, we use cv2.inRange() to look at our HSV frame and check it against our lower constraint (LC) and upper constraint (UC). This generates a Mask—a pure black-and-white image where pixels within the target color space are completely white (255), and everything else is completely black (0).

By taking that mask and running a fast bitwise operation

We force the computer to evaluate every single pixel. If the mask is zero, the output is blacked out. If the mask is active, the original, rich color information passes through perfectly, isolating our target object from the background noise instantly.

Get your circuits wired up, get this script running on your machine, and let me know in the comments section below what kind of performance numbers you are pulling on your local workbench. I’ll catch you guys in the next lesson!

Remember we are still setting the LED color to the color that cursor is pointing at. This is the circuit for connecting the RGB LED.

Fusion Hat Circuit Diagram
This is the circuit we will use moving forward in the class
Uncategorized

NVIDIA Jetson Orin Nano: How to Create and Use Swap Space for Larger Local LLM Models

In our first LLM class, we learned that we could run several of the smaller LLM models on the Jetson Orin, and then in the last class, we saw that we could make those small models run on the GPU, and we did notice faster performance as we began to move the workload to the GPU. The next challenge we found was that the larger models could give EOF errors, which were End of File errors, which usually means we have crashed due to not enough memory.  So, we need to work through this more methodically, and we need to run real benchmarks. Remember, we are running with the Big Dogs, and we are finding that we face tradeoffs between running on the CPU more slowly, or switching to the GPU and facing unpredictable throttling.

Our approach today is to deal with the memory issue. We will begin by turning off the GPU modifications, and just operate on the CPU. We will address the memory issue by creating swap space, and then we will benchmark our models running on the CPU, and complete a spreadsheet. We will need to begin by removing our configuration file that pointed us to the GPU:

Now lets create some swap space by allowing the system to use the SD card or the NVME for memory. Note, I am booting on a NVME. If you are using a SC card, you will notice more performance degradation by using swap space.

This command will create a swapfile:

This command will activate the swapfile:

This command will turnoff use of the swap space

This command shows you if the swapfile is being used

Now lets activate the swapfile:

Now each of these models should run on our system, as the swapfile will prevent the EOF error. Larger models will take a hit because they are bigger, and hence run slower, and then they bigger models will take a further hit in speed because swap memory will be slower than system RAM,

 

Model

 Model Family Size / Parameter Count Best Used For
gemma3:1b Google Gemma 3 1 Billion Ultra-fast responses, light footprint
llama3.2:1b Meta Llama 3.2 1 Billion High-efficiency conversational loops
phi4-mini:3.8b Microsoft Phi-4 3.8 Billion Heavy reasoning and coding logic
qwen3:4b Alibaba Qwen 3 4 Billion Structured data and multilingual logic
qwen3.5:4b Alibaba Qwen 3.5 4 Billion Advanced context processing
gemma3:4b Google Gemma 3 4 Billion Maximum analytical depth on Orin Nano
nemotron-3-nano:4b NVIDIA Nemotron 3 4 Billion Edge-optimized reasoning and tool-use

For this lesson, we will just be using CPU computation. This will allow us to benchmark simply between models.

Benchmarking Local LLM on NVIDIA Jetson Orin on Jetpack 7.2        
Model CPU/GPU Power Prompt Rate Eval. Rate Throttling Correct Answer Swap
gemma3:1b CPU MaxN
gemma3:1b GPU MaxN
gemma3:1b GPU 25 Watts
gemma3:1b GPU 15 Watts
llama3.2:1b CPU MaxN
llama3.2:1b GPU MaxN
llama3.2:1b GPU 25 Watts
llama3.2:1b GPU 15 Watts
phi4-mini:3.8b CPU MaxN
phi4-mini:3.8b GPU MaxN
phi4-mini:3.8b GPU 25 Watts
phi4-mini:3.8b GPU 15 Watts
qwen3:4b CPU MaxN
qwen3:4b GPU MaxN
qwen3:4b GPU 25 Watts
qwen3:4b GPU 15 Watts
qwen3.5:4b CPU MaxN
qwen3.5:4b GPU MaxN
qwen3.5:4b GPU 25 Watts
qwen3.5:4b GPU 15 Watts
gemma3:4b CPU MaxN
gemma3:4b GPU MaxN
gemma3:4b GPU 25 Watts
gemma3:4b GPU 15 Watts
nemotron-3-nano:4b CPU MaxN
nemotron-3-nano:4b GPU MaxN
nemotron-3-nano:4b GPU 25 Watts
nemotron-3-nano:4b GPU 15 Watts
NVIDIA

NVIDIA Jetson Orin Nano: Secret to Running Ollama on the GPU

One of the biggest frustrations with the new Jetpack 7.2 release is finding out that a standard installation of Ollama—the gold standard for running local LLMs—completely ignores your powerful NVIDIA GPU and defaults to the CPU.

In this lesson, we aren’t just going to fix that; we are going to measure the “truth” behind the performance. We will use data to see exactly how much gain we get from the GPU and where the hardware starts to hit the thermal throttling wall.

The Problem: The “Canned” Installation

When you run a standard Ollama install on the Jetson Orin Nano, the system doesn’t automatically recognize the integrated GPU (iGPU). If you open your NVIDIA Power GUI (jtop), you will see your CPU cores pegged at 100% while the GPU sits idle. This leads to slow response times and a disappointing experience.

Lets start by the standard ‘Canned’ Installation. The good news is, it is very simple:

To see exactly how your system is performing, run Ollama in verbose mode:

At this point you will have Ollama running a simple LLM locally on your Jetson Orin Nano. This is a huge step forward, but we now want to dig deeper and actually see how well this simple model is performing.  The first thing we do is run the Jetson Power GUI, hidden behind the NVIDIA icon in upper right of the menu bar.

Pay close attention to the Prompt Eval Rate and Eval Rate (tokens per second). These are our baseline numbers.

The “Secret Sauce” Solution

To force Ollama to use the Jetson’s CUDA cores, we have to manually override the system service configuration.

Step 1: Install the Nano Editor

Before we can edit system files, we need a reliable text editor. If you don’t have it yet, run this command:

Step 2: Create the Service Override

We need to tell the Ollama service exactly where to look for the GPU libraries. Use nano to open the following file:

Step 3: Add the Configuration

Copy and paste the following block into that file. This is the “Secret Sauce” that enables the iGPU and points the system to the correct CUDA backend:

Note: Save the file by pressing Ctrl+O, Enter, and then Ctrl+X to exit.

Step 4: Reboot

For the changes to take effect, We will do a reboot.

Benchmarking the Results

Once you have the GPU engaged, the real work begins. In the video, we look at a side-by-side comparison of performance across different Jetson Power Modes (10W, 15W, and MaxN).

Power Level Prompt Eval Rate (t/s)  Eval Rate (t/s) Throttling Observed?
CPU [Your Data] [Your Data] Yes/No
10W [Your Data] [Your Data] Yes/No
15W [Your Data] [Your Data] Yes/No
MaxN [Your Data] [Your Data] Yes/No

As we discovered, moving to the GPU provides a boost, but it also increases the heat signature. Watch the full video to see the charts and understand which power level provides the best “sweet spot” for stable, long-term AI performance on your Jetson Orin Nano. This is an important first step . . . getting the heavy lifting down to the GPU. Now in future videos we will explore how to get the work done Well on the GPU.

 

Making The World a Better Place One High Tech Project at a Time. Enjoy!