This Web-Based Proctoring System aims to provide a robust solution for remote exam monitoring using advanced computer vision and machine learning techniques. This system integrates multiple features to ensure the integrity of online assessments by analyzing various behavioral cues of test-takers.
Key components includes:
1. Head Pose estimation
2. Object Detection
3. Face Verification
4. Mouth Movement Analysis
5. Gaze Tracking
6. Face Spoofing
The model for Head Pose Estimation Was Created from the Scratch.
The Pandora dataset was used for the head posture module’s training and validation. There are 100 annotated sequences in the collection, gathered from 10 male and 10 female patients.Every participant gets five recordings. Regarding every topic,Two sequences with limited mobility are executed.adjusting the roll, pitch, and yaw angles independently.
The BIWI benchmark dataset was utilized to test the head pose module. It has fifteen thousand frames, with depth maps (640 × 480) and RGB (640 × 480). Twenty participants have been participated in the recordings: four of them were double-recorded, for twenty-four scenes in all. The actual values of pitch, yaw, and Along with the head center and the roll angles, the matrix of calibration.
OpenCV’s YOLO object detector to identify instances of banned items such as mobile phones, laptops, TVs, and books.
For Face Verification Dlib’s face verification model was used to identify the examinee by name. The model leverages a pre-trained ResNet50 CNN to generate a 128-dimensional feature vector for each facial image in the database.
To determine if an examinee is real or using a photograph, a face spoofing detection function ality was implemented. The process begins by capturing the examinee’s face image through the Face Detection module. This image is then converted into YCrCb and CIE Luv* color spaces using OpenCV. Histograms are computed for both color spaces and concatenated. The concatenated histogram is then fed into Scikit-learn’s ExtraTreesClassifier model to classify the face as either real or spoof.
Dlib’s pre-trained network was used to detect and predict 68 facial landmarks on the user’s face.
The lip region is identified using landmarks 60 through 67. To assess whether the mouth is open or closed, Lip Aspect Ratio (L.A.R) is defined as:
L.A.R = |P(62)−P(66)| |P(60) − P(64)|
where P(i) denotes the position of landmark i. Through extensive testing, threshold is fixed to 0.1. If the L.A.R exceeds 0.1, it indicates that the mouth is open; otherwise, it is considered closed. Additionally, to determine if the user is speaking, a buffer mechanism is employed. If the mouth remains open for more than 10 consecutive frames, it is classified as speaking,
