Abstract

Candidate Bib recognition in Physical Endurance Tests is streamlined with Image Processing. With a motive to integrate rational methods in the case of these tests, AI has a favorable impact. A system to detect and recognize individuals with real-time data setup supporting high performance and speed gives immediate responses.

Problem

Identifying the cases of impersonation in Physical tests of civil force recruitment exams can be manually tough. Any candidate can get substituted by another within seconds which might not be caught in the act. With this, the result processing gets chaotic. 

Solution

We introduced a system that functions with AI and performs Bib detection and recognition on each candidate’s linked Bib number. Our process includes capturing shots at the training scene and recognizing Bib placed around the candidates’ chest. Firstly, the detection process is carried out by identifying the Bib number and confirming whether the number tag is worn by the same individual throughout the test.

Then the recognition process is carried out by determining whether the assigned Bib number is worn by that registered individual by referring to our previously collected candidate registration database.

Functions:

1. Dataset collection: A set of candidate videos from the PET event are collected and converted into images. The videos must be crystal-clear to detect the numbers easily.

2. Data annotation: We deployed the LabelImg tool to label the necessary objects while training the system. In this case, we first annotate the person from the image and then the Bib number attached to them.

The annotated objects are stored in an XML file and are reinspected with Python script.

3. Data splitting:The annotated data is further split into train, validation, and test data in required percentages. This data is converted and stored in TensorFlow records, consuming less space to store and transfer suitable quantities of data.

4. Data training: We train the model for object detection with TensorFlow 2. Applied SSD MobileNet v2 model supporting high performance and speed. The evaluation can be accessed during the training to view the process.

5.Monitoring the progress with TensorBoard:Through TensorBoard, a feature named TensorFlow is used such that it helps us monitor the training task with ease. Once minimal to no loss is reported after we examine the training loss, validation loss, and accuracy, we export the model and test the test data.

6. Data testing: Testing is performed on the exported model to produce the final product.

Tools and Software used:

LabelImg

Machine Learning