Hardware-Software Co-development for Audio and Video Data Acquisition and Analysis
György Kalmár · 2021
similar' samples.The data may be diverse and pattern recognition includes many different fields and applications, such as signal processing, image analysis, or speech recognition, to name a few.In the past, pattern recognition involved algorithms designed by experts.Later, machine learning procedures emerged to solve the classification section of the problem.These methods were trained on compressed and irredundant input, called feature vectors.These features were hand-crafted and required prior knowledge related to the specific areas.Today, with the advent of artificial intelligence (AI), pattern recognition is dominated by AI algorithms, and the term itself usually implies that some sort of machine learning method is involved.These AI procedures take the training dataset, perform automated feature extraction and selection, on which the classification task is optimized.This scenario is called supervised learning, which requires a (potentially large) labeled dataset during the training process.Its counterpart, unsupervised learning, such as clustering, processes training data that has not been hand-labeled, and attempts to find inherent patterns in the data that can later be adopted to determine the correct label for new data instances.In the current work, only supervised methods are discussed ranging from classical solutions to modern, AI-based algorithms.The key component of any data analysis procedure is data it is trained and tested on.Mainly, quality and quantity are significant.To explain this, the reader can visualize, for example, the classification problem as the separation of an N-dimensional space into M disjoint subspaces, where N is the length (dimension) of the data (signal, image, audio, text, feature vector, etc.) and M is the number of classes.The labeled data points fill discrete points of this space.The task is to separate the labeled 1 Chapter 2 Animal-Borne Anti-Poaching SystemAcoustic gunshot detection has been an area of active research in recent decades.Many algorithms and systems were developed to realize a low-cost, low-power, and reliable solution.In this chapter, a wearable, animal-borne gunshot detector is introduced, which offers ultra-low power consumption and enhanced detection accuracy.The device is integrated into consumer GPS tracking collars, therefore, the combined system is able to send GPS-tagged gunshot alerts to support law enforcement.The main novelties of the work are a specially designed wake-up procedure that allows low-power constant listening and a two-domain based gunshot detection algorithm.Hardware-level modifications solved the critical power-consumption problem and provided more information for the detector, which resulted in enhanced classification accuracy.The structure of the chapter is as follows.Section 2.1 and Section 2.2 introduce the problem and list related research and notions.The succeeding sections explain our approach and system design, and detail the novel wake-up mechanism and compare it to traditional solutions.In Sections 2.7 and 2.8, the architecture of the developed system is presented.In Section 2.9, the evaluation of the system and its results are included.Section 2.10 presents a brief exploration to possible data-driven approaches.In the closing section, the chapter is summarized and final thoughts are aggregated.