Preliminary results of a proposed CNN framework for use in motorway applicable detection systems

This paper presents the preliminary results of the first stage of a proposed framework of existing image detection convolutional algorithms, and sets the image portable detector requirements, focusing on portable detection devices. Such devices perform image segmentation and transmit monitoring meta-data. Four pre-trained CNNs have been used to classify images of vehicles passing through toll posts, and the results are presented and compared. At subsequent stages of the presented framework certain types of vehicles will be detected and categorized, based on the type of cargo they carry. The proposed framework is wrapped by an appropriate system architecture which interacts with incident response or facility management systems, and tries to achieve precise speed/memory/accuracy balance for each targeted detection application accordingly. To this end, various ways are investigated and evaluated, to trade accuracy for speed and memory usage in modern mobile convolutional algorithm driven detection systems monitoring specific regions of interest. Such tradeoffs provide increased detection capabilities to incident response systems that interact with such detectors in real-time, or in post-processing with facility management systems for accurate and automated decisions adaptation.