CORTEXA
← Browse
zenodoJournal article2026-07-28

Radar-Based Multi-Class Recognition of Pedestrians, Cyclists, and Vehicles Using Random Forest and LSTM Under Clutter

Md. Abdur Rouf

Reliable radar-based discrimination of pedestrians, cyclists, and vehicles is a core requirement for roadside safety monitoring and autonomous-driving perception, particularly under adverse lighting and weather where camera-only systems degrade. This paper presents a machine-learning framework for three-class radar target recognition human (pedestrian), car, and bicycle evaluated under both an ideal, clutter-free environment and a realistic environment incorporating ground, weather, and urban clutter together with additive noise. A synthetic phased-array radar dataset (9.4 GHz, 500 samples per class per condition) was generated from the monostatic radar range equation, target-specific radar-cross-section and micro-Doppler models, and standard ground/weather/urban clutter distributions. Two classifiers were trained and compared: a Random Forest (RF) operating on aggregated per-dwell handcrafted features, and a Long Short-Term Memory (LSTM) network operating directly on the raw per-pulse feature sequence. Under ideal conditions, RF achieved 98.1% accuracy (F1 = 0.981, AUC = 0.9997) and LSTM achieved 89.3% accuracy (F1 = 0.892, AUC = 0.983). Under realistic clutter, RF accuracy changed only marginally, to 97.9%, while LSTM accuracy fell to 87.2%. After mutual-information-based feature selection and hyperparameter tuning, RF remained at 97.9% (already near ceiling) while tuned LSTM recovered to 89.1% (F1 = 0.891, AUC = 0.972). These results indicate that classifiers built on statistics aggregated over a full radar dwell are substantially more robust to per-pulse clutter than sequence models that consume raw per-pulse observations directly, and that hyperparameter tuning recovers a meaningful share of the accuracy lost by the latter. All results reported are from an actually executed simulation and training pipeline; validation against measured radar data is identified as necessary future work.

View free PDFSource page

Related papers

zenodoJournal article2026-08-01

Análisis técnico-jurídico de las posibilidades de la Artillería Antiaérea ante la Amenaza de los sUAS

José Manuel Castro Milla

LEGAL REVIEW Castro Milla, José Manuel. “Análisis técnico-jurídico de las posibilidades de la Artillería Antiaérea ante la Amenaza de los sUAS.” Boletín CODESEL, vol. 2, no. 10, August 2026, ISSN-e: 3045-7750. Review Fe…

View free PDFSource page
zenodoJournal article2026-07-29

A Survey Paper on Rural E-marketplace for Local Farmers with Chatbot for Pesticides

Likhith Reddy D, Prajwal Mallappa Sankanur, Kushal K, Dr. Deepak N R

Rural farmers encounter issues regarding market access, crop advisory services, and safe use of pesticides, which results in lower income and productivity. The authors propose the design and evaluation of a smart digital platform for a rural emarketplace with an integrated AI-pow…

View free PDFSource page
zenodoJournal article2026-07-29

AI-Powered Student Mental Health Analytics and Academic Prediction System: A Supervised Machine Learning and Explainable AI Approach

Sagara C P, Mohammed Zaid, Dr T. Vasudev

Student mental health difficulties such as stress, anxiety, depression and poor sleep frequently go unnoticed until they have already affected attendance, grades and personal wellbeing, because traditional counselling depends on a student voluntarily seeking help. This paper pres…

View free PDFSource page
zenodoJournal article2026-07-29

Design and Block-Level Simulation of a 16-Bit SAR ADC in 180 nm CMOS for GPR Applications

Aarushi Kulkarni, B Balaaditya, Dr. Kiran Bailey

This paper presents the design and block-level simulation of a 16-bit Successive Approximation Register (SAR) Analog-to-Digital Converter (ADC) implemented in the 180 nm CMOS Process Design Kit (PDK) of the Semiconductor Laboratory (SCL), India, using Cadence Virtuoso with the Sp…

View free PDFSource page