Detection of people from above with Thermal Camera

edit

I have a MLX90640 thermal camera 13' above the ground. it has 32x24 resolution with a 110 x 75 Field of View.

I am trying to narrow the temperature range from -40 - 300 to a better range of 6 - 20. i think I am doing this right but am not sure.

I am trying to determine if contour detection or blob detection would be better for counting the number of people in the room. I am new to OpenCV and am not sure if I am using the image filters in the best way.

I posted the code below. Any advice or recommendations are appreciated.

Image of just one person.

The sensor sometimes gives back strange data like extra hotspots such as in the image below. It is still with the same one person but the temps have fluctuated. The blob detector counts 1 while the contours count 6.

import sys
import os
import time
import colorsys
import numpy as np
import cv2
import datetime

from PIL import Image

sys.path.insert(0, "./build/lib.linux-armv7l-3.5")

import MLX90640 as mlx

img = Image.new( 'L', (24,32), "black")

def irCounter():   


    mlx.setup(8) #set frame rate of MLX90640

    f = mlx.get_frame()

    mlx.cleanup()

    # get max and min temps from sensor
    v_min = min(f)
    v_max = max(f)

    # Console output for testing
    textTime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S') # get timestamp
    print(textTime)
    print(min(f))
    print(max(f))
    print("")

    for x in range(24):
        row = []
        for y in range(32):
            val = f[32 * (23-x) + y]
            row.append(val)
            img.putpixel((x, y), (int(val)))

    # convert raw temp data to numpy array
    imgIR = np.array(img)

    ## Threshold the -40C to 300 C temps to a more human range
    # Sensor seems to read a bit cold, calibrate in final setting
    rangeMin = 6 # low threshold temp in C
    rangeMax = 20 # high threshold temp in C


    # Apply thresholds based on min and max ranges
    depth_scale_factor = 255.0 / (rangeMax-rangeMin)
    depth_scale_beta_factor = -rangeMin*255.0/(rangeMax-rangeMin)

    depth_uint8 = imgIR*depth_scale_factor+depth_scale_beta_factor
    depth_uint8[depth_uint8>255] = 255
    depth_uint8[depth_uint8<0] = 0
    depth_uint8 = depth_uint8.astype('uint8')

    # increase the 24x32 px image to 240x320px for ease of seeing
    bigIR = cv2.resize(depth_uint8, dsize=(240,320), interpolation=cv2.INTER_CUBIC)

    # Normalize the image
    normIR = cv2.normalize(bigIR, bigIR, 0, 255, cv2.NORM_MINMAX, cv2.CV_8U)

    # Apply a color heat map
    colorIR = cv2.applyColorMap(normIR, cv2.COLORMAP_JET)

    # Use a bilateral filter to blur while hopefully retaining edges
    brightBlurIR = cv2.bilateralFilter(normIR,9,150,150)

    # Threshold the image to black and white 
    retval, threshIR = cv2.threshold(brightBlurIR, 210, 255, cv2.THRESH_BINARY)

    # Define kernal for erosion and dilation and closing operations
    kernel = np.ones((5,5),np.uint8)

    erosionIR = cv2.erode(threshIR,kernel,iterations = 1)

    dilationIR = cv2.dilate(erosionIR,kernel,iterations = 1)

    closingIR = cv2.morphologyEx(dilationIR, cv2.MORPH_CLOSE, kernel)

    # Detect edges with Canny detection, currently only for visual testing not counting
    edgesIR = cv2.Canny(closingIR,50,70, L2gradient=True)

    # Detect countours
    contours, hierarchy = cv2.findContours(closingIR, cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE)

    # Get the number of contours ( contours count ...