Second Sight: MobileNet v1 Integration in Dynamic and Time Critical Scenarios

When contending with the safety and awareness of active shooters or one who may be carrying an instrument of destruction, a solution exists which may enable others to quickly respond and become aware of the presence of a firearm. For additional insight and spatial awareness of potentially dangerous activities, machine learning and computer vision further augment the span of abilities to increase one's sense when navigating through an environment. Convolutional Neural Network (CNN) architectures provide incredible accuracy in object recognition when trained unto a specific subject domain through feature extraction. However, CNNs may become computationally expensive and require indefinite amounts of hardware to support real-time vision propagation -which imposes difficulty if one was adorned with computer vision equipment such as body cameras. MobileNet is a computer vision paradigm specifically designed to perform on compact computational devices while retaining object recognition performance. This paper proposes the analysis and implementation of MobileNet in low computational capacity devices -such as body cameras-and still enacts high object recognition performance in dynamic and time-critical scenarios to further augment an officer's insight into potentially dangerous situations. A custom firearm dataset was produced of over 10,000 images and the model performed over 85% on weapon validation accuracy when provided and trained on a variety of different dynamic and active environments containing the presence of firearms.