vehicle_counter

Detects and counts vehicles moving from left to right or vice versa. Also for raspberry pi. Includes speed approximation.

Procedure

Process the frame

• Convert original frame to grayscale and remove noise
• Create an average frame, which is basically the static background. Each new frame impacts this average frame to some extent
• Compute the difference between average and current frame to get the foreground
• Apply a threshold to this difference frame to get clean shapes of the moving objects in the foreground

Get the contours

• Get the contours and filter out small ones
• Reduce the amount of points each contour consists of and sort the contours by size
• Find contours which are close to each other and merge these. For instance, if a is close to b and b is close to c, then a is also close to c via b. Consequently, the new merged contour consists of all three contours a, b and c

Get the centroids and add these to vehicles or create new vehicles

• Vehicles are stored as dictionaries with the attributes: id, fist time seen, last time seen, direction, a boolean 'found' and a list of 'tracked' centroids (blue dots)
• If the distance isn't too great and the direction is correct, the centroid will be added to this vehicle. An added centroid will be removed from the list of candidates
• If there are still centroids left in the candidates list, check if a new vehicle can be created
• If a centroid from the current frame is close to one of the last frame, a new vehicle will be created

Detect whether vehicles crossed center line

• Vehicles which haven't been seen for some time will be removed and its data saved to a .csv-file
• If a vehicle has not been found yet and has crossed the center line, the vehicle counter increments
• If a vehicle has crossed the left and right barrier, its speed will be calculated

Materials

• Speedtrack repo by @iandees
• Tutorial about contours
• Blog post about raspberry pi threading on pyimagesearch by Adrian Rosebrock