There can be many for example:

1. These cameras or the supporting systems are not designed for large computational and memory-based processing.

2. I am a bit confused about your question, it can work well for face detection i.e. implementing an SSD or YOLO on CCTV, but for recognition, you need training and training datasets.

Khurram Hameed Yes, face detection is possible and various cctv cameras are also present for face detection. I just want some information on the factors that depends on face recogntition in cctv cameras. What are they? And if possible, how they can be tackled?

Hi Sahil,

A security-surveillance integrator I worked for conducted field tests with multiple (i) face recognition algorithms, (ii) IP camera make/models, (iii) compute platforms on which to run these algorithms.

Regardless of the specific algo-cam-platform combo, here are some of the major problems we observed:

0. In any real world deployment scenario the makes/brands of the cameras already installed will vary. This can be a very serious problem. Same thing with the NVRs already installed on the field. Finally, time-stamps. This is not specific to face recognition. Dealing with the variations/errors in the frame time-stamps of the video streams can be a ---HUGE--- headache. Especially if you have unskilled and under-educated engineering teams that develop or deploy those cameras and NVRs...

1. Camera height. If the installation height is above 4 meters it is pretty much impossible to get satisfactory performance (especially in terms of missed detections)

2. Illumination and reverse light. IR cameras may help with this but they have their own problems.

3. Resolution. Depending on the target distance your so called face image can be 25 - 30 pixels... In case you have a full body description you can partially remedy this by combining body re-id with face re-id, but do not expect miracles.

4. Hoodies, baseballs hats, face masks !!!! , hijabs...

5. It may not be obvious at first but matching detected faces with your face database becomes an issue as your database hits about 90-100 million IDs (each with multiple face pictures).

6. Result reporting and logging. If/when you reach an acceptable perf level, you will deploy the system and turn it on. Depending on your use case, results will start pouring in at some point. You will need good and user friendly software to make those results useful. Camera operators are not rocket scientists...

---good luck.