International Journal of Electrical and Computer Engineering (IJECE) Vol. 14, No. 3, June 2024, pp. 3187~3196 ISSN: 2088-8708, DOI: 10.11591/ijece.v14i3.pp3187-3196  3187 Journal homepage: http://ijece.iaescore.com Deep learning and quantization for accurate and efficient multi- target radar inference of moving targets Nyasha Ernest Mashanda 1 , Neil Watson 1 , Robert Berndt 2 , Mohammed Yunus Abdul Gaffar 3 1 Department of Statistical Sciences, Faculty of Science, University of Cape Town, Cape Town, South Africa 2 Radar and Electronic Warfare, Defence and Security, Council for Scientific and Industrial Research, Pretoria, South Africa 3 Department of Electrical Engineering, Faculty of Engineering and the Built Environment, University of Cape Town, Cape Town, South Africa Article Info ABSTRACT Article history: Received Jun 14, 2023 Revised Sep 3, 2023 Accepted Sep 12, 2023 Real-time, radar-based human activity and target classification is useful for wide-area ground surveillance. However, the feasibility of deploying deep learning (DL) models in radar-based systems with limited computational resources remains unexplored. This paper investigated the effect of quantization on model throughput and accuracy for deployment in radar systems. A seven-layer residual network was proposed to classify ground- moving targets and achieved a test accuracy of 87.72%. The model was then quantized to 16-bit and 8-bit precision, resulting in a 3.8 times speedup in inference throughput, with less than a 0.4% drop in test and validation accuracy. The results showed that quantization can improve inference throughput with a negligible decrease in target classification accuracy. The increase in throughput and reduction in computational expense that comes with quantization promotes the feasibility of the deployment of DL models in systems with limited computational resources. The findings of this paper hold significant promise for the successful use of quantized models in modern radar systems, while adhering to stringent size, weight and power consumption constraints. Keywords: Convolutional neural network Inference Micro-Doppler Quantization Radar This is an open access article under the CC BY-SA license. Corresponding Author: Nyasha Ernest Mashanda Department of Statistical Sciences, Faculty of Science, University of Cape Town Rondebosch, Cape Town, South Africa Email: nyashamash001@gmail.com 1. INTRODUCTION Human activity and target classification has gained interest in recent years due to its applications in indoor and outdoor surveillance systems for health monitoring [1], border control [2] and security [3]. Different sensor systems have been used for classification, including cameras [4], Lidar [5] and radar [6]. Unlike cameras or Lidar, the performance of a radar sensor is less sensitive to varying weather conditions and different levels of light [7]. Furthermore, radar offers a more extended detection range than optical sensors [6] and can detect targets behind opaque objects [8]. These advantages make radar more suitable for outdoor surveillance systems to curb illegal activities such as poaching, smuggling and livestock theft. Traditionally researchers have used various techniques to extract pre-defined features for classification from radar data. Examples of such pre-defined features include those related to the physical characteristics of the target [9] and discrete cosine transform coefficients [10]. The features were then used for classification using support vector machines [11] or random forests [12]. This feature estimation and extraction process is highly dependent on human experience and domain knowledge, which makes it susceptible to human error. The advent of deep learning (DL) has allowed an alternative approach to solving