Strathmore Research Brown Bag Sessions

Permanent URI for this community

http://hdl.handle.net/11071/6431

Browse

Browsing Strathmore Research Brown Bag Sessions by Subject "Computer Vision Techniques"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Concealed firearm detection on video surveillance using skeletal tracking and computer vision techniques
    (Strathmore University, 2019-01-30) Muchiri, Henry
    Crimes involving the use of illegal firearms have been on the rise worldwide. These illegal firearms are carried in a concealed manner under a person’s clothing and specifically on the hip area for easy access. To counter this threat, security providers have adopted various detection techniques such as metal detectors, image sensors and behavioural analysis by trained officers. These techniques, however, are primarily oriented to detecting metal, require the person being screened to be stationary and cooperative and are employed at entrances of enclosed facilities such as airports, buildings, and shopping malls, among others. There remains a need for a concealed firearm detector which can detect firearms remotely, in open places such as streets and walkways as people go about their business. This study leverages computer vision and human motion tracking techniques to develop a machine learning model with the ability to distinguish between armed and unarmed subjects based on their motion characteristics. This is informed by studies on behavioural analysis of persons carrying concealed firearms which have found motion as a key indicator. Simulated videos of 26 individuals walking with a firearm concealed on the hip were recorded using a time of flight depth camera. The recorded data consisted of 3D spatial-temporal depth data of tracked joints on the human body. Relevant features were extracted from the data and used to train various machine learning models which include Support Vector Machines (SVM), Naïve Bayes, K-Nearest Neighbour, J48 and Random Forest. Experiments were conducted to tune and evaluate the model’s performance. Preliminary results are impressive with the K-nearest neighbours algorithm producing the best results of 93.11% correct classification rate, 86.22% Kappa coefficient and under 7% false positive rate. These results indicate the viability of the approach to detect concealed firearms remotely while persons are in motion. This approach can be integrated into existing CCTV networks on streets and walkways, hence enabling the remote screening of people while they move about doing their business. In addition, remote screening will ensure the safety of security personnel who presently must be in close proximity with those they are screening.