Integrating IMU accelerometer data to estimate position

I am trying to estimate position change using an MPU6050 IMU. I know that drift is a problem, but I am only trying to do this over short periods of less than a few seconds.

I've adjusted for gravity using sensor fusion madgwick filter and after adjustment my acceleration readings look good, but when I double integrate them, the resulting positions are very very small.

Any insight into what I might be doing wrong?

VectorFloat getAccelerationNoGravity() {
  frame++;
  auto timeSinceLastSample = ReadTime(); 

  /**
   * Get the current acceleration readings compensated for gravity
   * 1. Get raw readings from acc & gyro
   * 2. Feed into madgwick filter, get quaternion
   * 3. Get gravity vector from quaternion
   * 4. subtract gravity from raw acceleration readings
   */

  // 1. Get raw readings from acc & gyro
  sensors_event_t acc, gyro, temp;
  mpu.getEvent(&acc, &gyro, &temp);

  // 2. Feed into madgwick filter, get quaternion
  float x = acc.acceleration.x;
  float y = acc.acceleration.y;
  float z = acc.acceleration.z;
  float gx = gyro.gyro.x;
  float gy = gyro.gyro.y;
  float gz = gyro.gyro.z;
  float deltat = timeSinceLastSample / 1000000.0;

  madgwickQuaternionUpdate(q, x, y, z, gx, gy, gz, deltat);
  Quaternion quat = Quaternion(q[0], q[1], q[2], q[3]);

  // 3. Get gravity vector from quaternion
  VectorFloat gravity = getGravity(&quat); // returns percentages of gravity

  // 4. subtract gravity from raw acceleration readings
  VectorFloat accAdj = getLinearAcceleration(VectorFloat(x, y, z), gravity);

  // 5. calculate bias & adjust
  calibrateBias(accAdj); // average of x samples
  accAdj.x -= bias.x;
  accAdj.y -= bias.y;
  accAdj.z -= bias.z;

  // calc position
  auto accMagnitude = accAdj.getMagnitude();
  if (biasComputed /*&& accMagnitude > 0.1*/) {
    // v0 (initial velocity) = v
    auto v0x = currVel.x;
    auto v0y = currVel.y;
    auto v0z = currVel.z;

    // currVel (current velocity) = v0 + a * t
    currVel.x = v0x + deltat * accAdj.x;
    currVel.y = v0y + deltat * accAdj.y;
    currVel.z = v0z + deltat * accAdj.z;

    // delta_x = v0 * t + 1/2 * a * t^2 * 100 (m -> cm)
    float deltat_sq = deltat * deltat;
    currPos.x += v0x * deltat + 0.5 * accAdj.x * deltat_sq * 100;
    currPos.y += v0y * deltat + 0.5 * accAdj.y * deltat_sq * 100;
    currPos.z += v0z * deltat + 0.5 * accAdj.z * deltat_sq * 100;
  }

  t_compute = micros() - t_compute;
  if (biasComputed && frame % 50 == 0) {
    Serial.printf("%f,%f,%f\n", accAdj.x, accAdj.y, accAdj.z);
  }

  return accAdj;
}

This results in readings like this

0.117377,0.135253,0.010953
0.117308,0.133007,0.010974
0.117446,0.129459,0.010982
0.117550,0.125331,0.010972
0.117732,0.120880,0.010971
0.117961,0.115567,0.011018
0.118101,0.111308,0.011070
0.118161,0.112330,0.011132
0.118256,0.114275,0.011229
0.118401,0.116992,0.011322

but I am moving the IMU more than 50 cm so it seems the result is orders of magnitude off.

Again, my adjusted accelerometer readings make sense so perhaps I made a mistake implementing the kinematics?

I don't have enough reputation to comment that's why writing as an answer but this is actually a comment. I think your deltat should be 1/sampling_rate (ex: 1/100 if you get 100Hz data).