Key aspects of the project:

• Motion detection:

The system continuously monitors the field of view of the ESP32-CAM module and flags a finded face.

• Laser control:

Using two servomotors, a laser dot is automatically aimed at the detected target.

• Purpose:

Demonstrating the integration of a simple machine vision algorithm on a microcontroller with limited resources and real-time control of the laser’s position.

• Motivation:

Building an accessible, low-cost security solution using off-the-shelf components.

• Applicability:

1. Educational prototype in robotics
2. Low-cost security system

Block Diagram of the system:

Data and signal flow:

• Power supply (5 - 12V) → 5V regulator → ESP32-CAM
• ESP32-CAM: image processing → X/Y coordinate calculation → PWM signal
• Servo 1 (pan) and Servo 2 (tilt): orient the laser mount
• Laser: tracks the target by projecting a dot

Component List

Electrical schematic:

The hardware part is completely soldered

Development Environment

• Arduino IDE 2.0 with ESP32 support (Espressif plugin)
• Board: AI Thinker ESP32-CAM

3rd-party Libraries

• Espressif ESP32 Camera: esp_camera.h and img_converters.h for camera init and JPEG/RGB565 conversion
• ArduinoWebsockets: WebSocketsServer.h for real-time frame streaming over WebSockets
• Arduino Core WiFi & HTTP: WiFi.h and WebServer.h for AP connection and HTTP control endpoint
• ESP32Servo: for precise PWM servo control
• Framebuffer Graphics: fb_gfx.h for drawing bounding boxes on RGB565 frames
• Espressif ML Face Detection: human_face_detect_msr01.hpp and human_face_detect_mnp01.hpp for two-stage face detection

Face detection algorithm Face detection is implemented in the FaceFinder class:

• Candidate generation with HumanFaceDetectMSR01.
• Refinement with HumanFaceDetectMNP01.
• If any faces remain, the first prediction’s bounding box (x, y, width, height) is extracted and attached to the frame’s dimensions.
• Optionally the box is drawn (or filled) directly into the frame buffer using fb_gfx_drawFastHLine, fb_gfx_drawFastVLine, or fb_gfx_fillRect.
• A boolean found flag is set to true, and the coordinates stored in face.x, face.y, face.w, face.h.

    bool find(uint8_t *buf565, uint16_t width, uint16_t height, bool draw = true, bool fill = false) {
      found = 0;
      frame_w = width;
      frame_h = height;
      {
        HumanFaceDetectMSR01 s1(0.1F, 0.5F, 2, 0.3F);
        HumanFaceDetectMNP01 s2(0.4F, 0.3F, 1);
        std::list<dl::detect::result_t> &candidates = s1.infer((uint16_t *)buf565, {height, width, 3});
        std::list<dl::detect::result_t> &results = s2.infer((uint16_t *)buf565, {height, width, 3}, candidates);
 
        if (!results.size())
          return 0;
 
        std::list<dl::detect::result_t>::iterator prediction = results.begin();
        x = (int)prediction->box[0];
        y = (int)prediction->box[1];
        w = (int)prediction->box[2] - x + 1;
        h = (int)prediction->box[3] - y + 1;
        if ((x + w) > width)
          w = width - x;
        if ((y + h) > height)
          h = height - y;
        results.end();
      }
 
      if (draw) {
        fb_data_t fbd;
        fbd.width = width;
        fbd.height = height;
        fbd.data = buf565;
        fbd.bytes_per_pixel = 2;
        fbd.format = FB_RGB565;
        uint32_t color = 0b1111100000000000;
        if (fill) {
          fb_gfx_fillRect(&fbd, x, y, w, h, color);
        } else {
          fb_gfx_drawFastHLine(&fbd, x, y, w, color);
          fb_gfx_drawFastHLine(&fbd, x, y + h - 1, w, color);
          fb_gfx_drawFastVLine(&fbd, x, y, h, color);
          fb_gfx_drawFastVLine(&fbd, x + w - 1, y, h, color);
        }
      }
      found = 1;
      return 1;
    }

Servo and laser actions are done on the core0:

1. Initialization (initServos()):

• Set PWM period to 50 Hz for both servos.
• Attach X and Y servos on GPIO 13 and 12 with calibration pulses (MIN_PULSE, MAX_PULSE).
• Center servos at midpoints (X_MAX/2, Y_MAX/2).
• Configure laser pin (GPIO 2) as OUTPUT and default LOW.

void initServos() {
  servoX.setPeriodHertz(50);
  servoY.setPeriodHertz(50);
  servoX.attach(SERVO_X_PIN, MIN_PULSE, MAX_PULSE);
  servoY.attach(SERVO_Y_PIN, MIN_PULSE, MAX_PULSE);
 
  servoX.writeMicroseconds(map(lastCamX, 0, X_MAX, MIN_PULSE, MAX_PULSE));
  servoY.writeMicroseconds(map(lastCamY, 0, Y_MAX, MIN_PULSE, MAX_PULSE));
}

2. Automatic tracking (trackTargetX, trackTargetY):

• Constrain camera coordinates to [0, X_MAX] or [0, Y_MAX].
• Apply a dead-zone threshold (X_THRESHOLD, Y_THRESHOLD) to avoid parasite movements.
• Map camera X/Y to microsecond pulse widths (inverted so left/right or up/down correspond correctly).
• Update servos via writeMicroseconds().

// function to move the pan servo
void trackTargetX(int camX) {
  camX = constrain(camX, 0, X_MAX);
  if (abs(camX - lastCamX) < X_THRESHOLD)
    return;
  lastCamX = camX;
  servoX.writeMicroseconds(map(camX, 0, X_MAX, MAX_PULSE, MIN_PULSE));
}
 
// fucntion to move the tilt servo
void trackTargetY(int camY) {
  camY = constrain(camY, 0, Y_MAX);
  if (abs(camY - lastCamY) < Y_THRESHOLD)
    return;
  lastCamY = camY;
  servoY.writeMicroseconds(map(camY, 0, Y_MAX, MAX_PULSE, MIN_PULSE));
}

3. Manual control (when trackON == false):

• Flags left, right, up, down trigger fixed-time moves using SERVO_LEFT/SERVO_RIGHT or SERVO_UP/SERVO_DOWN for a duration, then reset to SERVO_STOP.
• shoot flag toggles the laser pin HIGH for a single loop iteration.

// example for moving to the left
if (left) {
        servoX.writeMicroseconds(SERVO_LEFT);
        delay(SERVO_X_MOVE_TIME);
        servoX.writeMicroseconds(SERVO_STOP);
        left = false;
 
      }

Setup

• Call cam_init(FRAMESIZE_HVGA, PIXFORMAT_JPEG, 10) to start the camera in JPEG mode.
• Connect to the configured WiFi (WIFI_SSID/WIFI_PASS).
• Start an HTTP server on port 80 to set control flags.
• Start a WebSockets server on port 82 for live frame broadcast.
• Launch the core0 task pinned to core 0 for servo and laser control.

void setup() {
  Serial.begin(115200);
  delay(200);
  cam_init(FRAMESIZE_HVGA, PIXFORMAT_JPEG, 10);
 
  WiFi.mode(WIFI_STA);
  WiFi.begin(AP_SSID, AP_PASS);
 
  while (WiFi.status() != WL_CONNECTED) {
    delay(500);
    Serial.print(".");
  }
 
  Serial.print("WIFI IP: ");
  Serial.println(WiFi.localIP());
 
  server.on("/action", HTTP_GET, handleAction);
  server.begin();
  Serial.println("HTTP server started on port 80");
 
  ws.begin();
  xTaskCreatePinnedToCore(core0, "Task0", 10000, NULL, 1, &Task0, 0);
}

Main Loop

• Call server.handleClient() and ws.loop() to process incoming HTTP and WebSocket events (e.g. /action?go=left).
• Grab a frame via esp_camera_fb_get(). If no frame, delay and retry.
• If trackON is true:
  1. Allocate an RGB565 buffer and convert the JPEG frame (jpg2rgb565).
  2. Run face.find() to detect and draw the bounding box.
  3. Re-encode to JPEG (fmt2jpg) and broadcast binary data over WebSockets.
• If trackON is false:
  • Broadcast the raw JPEG buffer directly.
• Return the frame buffer (esp_camera_fb_return) and delay ~20 ms to regulate frame rate.

void loop() {
  server.handleClient();
  ws.loop();
 
  // camera frame capture & send
  camera_fb_t *fbj = esp_camera_fb_get();
  if (!fbj) {
    delay(20);
    return;
  }
 
  if (trackON) {
    // Only convert image & run face-finder when tracking
    uint32_t len = fbj->width * fbj->height * 2;
    uint8_t *buf = (uint8_t *)ps_malloc(len);
    if (buf) {
      bool ok = jpg2rgb565(fbj->buf, fbj->len, buf, JPG_SCALE_NONE);
      if (ok) {
        // swap low->high byte
        for (uint32_t i = 0; i < len; i += 2) {
          uint8_t b = buf[i];
          buf[i] = buf[i + 1];
          buf[i + 1] = b;
        }
 
        // face detection
        face.find(buf, fbj->width, fbj->height, true, 0);
 
        // re-encode and broadcast
        if (ws.connectedClients()) {
          size_t jpg_buf_len = 0;
          uint8_t *jpg_buf = nullptr;
          ok = fmt2jpg(buf, len, fbj->width, fbj->height, PIXFORMAT_RGB565, 80, &jpg_buf, &jpg_buf_len);
          if (ok) ws.broadcastBIN(jpg_buf, jpg_buf_len);
          if (jpg_buf) free(jpg_buf);
        }
      }
      free(buf);
    }
  } else {
    // Just stream raw JPEG when not tracking
    if (ws.connectedClients()) {
      ws.broadcastBIN(fbj->buf, fbj->len);
    }
  }
 
  esp_camera_fb_return(fbj);
  delay(20);
}

At the time of the latest update, face detection in the frame occurs about twice per second (limited by the microcontroller’s processing power and the data transmission and processing speed). Of course, the project can be modernized and optimized to run as fast as possible! The end result is a fully functional project, ready for use, and its design can be customized to anyone’s preferences.

This project took me around 50–60 hours (maybe even much longer—after a while I stopped keeping track of time 😊), the hardest part being the software, where I ran into the most errors. The hardware side was fairly straightforward and didn’t give me any problems.

Archive with the project, unzip and open it in your Arduino IDE!

ESP32-CAM datasheet: https://media.digikey.com/pdf/Data%20Sheets/DFRobot%20PDFs/DFR0602_Web.pdf

ESP32Servo Library: https://github.com/madhephaestus/ESP32Servo

ESP32 Useful library and tutorials: https://randomnerdtutorials.com/

Espressif webpage: https://www.espressif.com/en/products/devkits/esp-eye/overview

Espressif Github: https://github.com/espressif/esp-who

Arduino and ESP32 Forums! And a lot of YouTube videos, but most of them was useless :)