Problem Statement
Face Recognition at Varying Angle
The main issue that this problem statement aims to solve is how to identify suspects by analysing CCTV feeds even when the frontal face is not clearly visible. To guard against theft and vandalism, CCTV cameras are frequently installed at various angles, usually at the top of poles, buildings, etc. As a result, taking a clear photo of someone’s face directly from the front is very challenging.
PS Number: PSAIML005
Domain Bucket: Artificial Intelligence
Category: Software
Dataset : NA
Create a face detection-based system that can recognise suspects in CCTV feeds even when just a portion of their features are visible due to the various positions that CCTV cameras are generally placed in. The solution must be flexible enough to accommodate persons with varying ages, skin tones, genders, and facial structures.
Background of the Problem
Identifying suspects by analysing CCTV feeds even when the frontal face has not been clearly captured is the key challenge this problem statement strives to overcome. CCTV cameras are typically fitted at varying angles, usually at the top of a pole, buildings etc. to protect it against vandalism and thefts. This makes it quite difficult to capture a clear picture of a person’s face directly from the front.
Objective
Build a face detection-based solution which can identify suspects in CCTV feeds even when their faces are partially visible due to the varying angles at which CCTV cameras are typically placed. The solution should be able to handle people of different age, skin colour, gender and facial structure.Solution should also maximize accuracy of object classification, object detection and object tracking. The solution should be designed in a highly scalable manner such that it is able to look for one or multiple persons across multiple videos concurrently. This is of relevance to large cities like Delhi or Mumbai where there are hundreds or thousands of CCTV cameras continuously capturing video footage every single second. It is understood that suitable hardware scale up may also need to be done. Solution should support both recorded videos as well as live CCTV feeds.
Summary
This is quite unlike ideal situations when a camera is conveniently placed closer to and right in the front (at the eye level) of a person, for example, getting a passport-sized photo clicked at a photo studio. The quality of image captured in both of the above situations is drastically different. Whereas there are several solutions based on facial recognition technology to search for a person in a video feed, the accuracy of results in case footage is captured under non-ideal conditions is generally poor. This is precisely the problem that the solution should overcome.