Automatic Passenger Counting in Busses - A Case Study

April 17, 2019

The following text is intended to give an overview of what modern image processing is capable of. The specific aim is to investigate the applicability of security cameras for the purpose of passenger counting using case studies.

In particular, different exposure conditions and utilizations are considered. Finally, advantages and existing challenges of the approach of using security cameras for passenger counting are analyzed. It should be noted that the pictures shown below did not come from security cameras but were taken by hand. Due to different perspectives, camera resolution and vehicle structure, the results may differ in real operation.

Scenario 1: Good exposure, normal utilization

image of a bus with people

Normal filled bus, in daylight. In total, 12 people can be recognized.

automatic people counting inside bus

All persons are detected correctly. Even the boy in the upper right corner and the man in the back left of the picture are recognized correctly although they are partially concealed. Furthermore, it should be emphasized that the girl in the middle right behind a glass pane is recognized despite reflection.

Detection rate: 12/12
Error: 0

Scenario 2: Good exposure, high utilization

image of a crowded bus with people

The same number of people as in the previous picture, however, in this image people walk in the middle.

image of a crowded bus with people - automatic people counting (APC) with Isarsense

Masking ensures that fewer people are recognized. In addition, a passer out of the bus is falsely recognized. It is interesting that the girl is recognized in the middle right, although only hair is visible. In addition to the pure person recognition additionally the handbag of the woman in the middle is recognized.

Recognition rate: 9/12
Error: 1 (person outside the bus)

Scenario 3: At night

image of a bus with people at night

Recording at night from behind. You can see 5 persons as well as the shoe of a sixth person.

automatic people counting in a bus at night with Isarsense

All people are corretly detected. It also worth pointing out that the person in the left corner is detected correctly, despite low contrast.

Detection rate: 6/6
Error: 0

Result overview

1 normal utilization12/12100%0
2 high utilization9/1275%1
3 at night6/6100%0

From the overview, it can be seen that the detection rate in the 3 scenarios considered does not vary significantly and is relatively high at 75% even in the worst case scenario. Furthermore, it can be seen that the error detections are generated only by persons outside the vehicle.

Advantages of the considered method

  • Relies on existing camera systems
    • No additional sensor hardware required in case cameras are already installed due to security concerns.
  • Detection of additional categories possible
    • Gender
    • Age
    • Wheelchairs
    • Strollers
    • Mobile phones
  • No accumulation of measurement errors
  • Fast calculation
    • 200 milliseconds on graphics card
    • 10 seconds on CPU
Back to blog overview