C-Mobile: A Lightweight Composite MobileNetV2 Model for Intrusive Object Detection under Power Grid Surveillance

Abstract

Intrusive object detection is a key task in real-time power grid surveillance, as the national smart grid is developing rapidly. It turns out to be time-consuming and inaccurate if the surveillance is manually performed by workers. Thus, with the booming of deep learning, we proposed an intrusive object detection algorithm, named C-Mobile, based on lightweight backbone MobileNetV2. To promote the interaction among features and ensure the real-time detection, we designed the composite MobileNetV2 backbone with an SE layer, where one of the MobileNetV2 can enhance the features of the other with a small increase in model complexity. To further utilize the extracted features, we proposed a top-down-bottom-up feature pyramid network (FPN) in which the bottom-up fusion with downsampling is applied after the traditional FPN and a cascaded region proposal network. Our dataset was collected through surveillance camera with 8,177 images and 17,883 object instances in five categories including trucks, cranes, lifts, excavators and pile drivers. Our C-Mobile reaches the highest mAP and the lowest model complexity on our dataset among state-of-the-art object detection algorithms, proving the efficacy of C-Mobile in real-time power grid surveillance.