Pedestrian recognition using micro Doppler effects of radar signals based on machine learning and multi-objective optimization

• Published in: Expert systems with applications 2019-12, Vol.136, p.304-315
• Main Authors: Severino, João Victor Bruneti, Zimmer, Alessandro, Brandmeier, Thomas, Freire, Roberto Zanetti
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Automobiles; Automotive radar; Autonomous car; Crashes; Embedded systems; Fatalities; Genetic algorithms; Human error; Kernel functions; Machine learning; Micro Doppler; Multi-objective optimization; Multiple objective analysis; Object recognition; Optimization; Optimization techniques; Pedestrian recognition; Pedestrians; Polynomials; Radar; Radar equipment; Signal quality; Support Vector Machine; Support vector machines; Traffic accidents; Velocity
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2019.06.048

•A radar sensor was used to detect pedestrians in near to crash accident situations.•Micro Doppler effects were better extracted by multi-objective optimization.•Datasets considering traffic accidents situations were captured by the radar.•Classifications involving four Support Vector Machine models where performed.•The proposed approach showed high accuracy in detecting pedestrians. Nearly 1.24 million people die every year in traffic accidents. More than 20% of the number of deaths are pedestrians. Traffic crashes will become, in 2030, the fifth leading cause of deaths worldwide. A considerable amount of these accidents occurs due to either imprudence or lack of dexterity of the driver. More than 70% of traffic accidents are attributed to human error, such as perception errors as distance judgment error or simply by lack of attention. Because of the problem aforementioned, this work proposes the development and discussion of a pedestrian recognition system using an automotive radar of 79 GHz. The main idea is to early detect pedestrians using the micro Doppler characteristics of the human body in near to crash situations (0–15 m). Based on these requirements, at first this work presents the improvement of velocity resolution of the radar system with the objective of extracting the micro Doppler characteristics of the objects. The improvement of velocity resolution was reached assuming the application of multi-objective optimization techniques, in this case Genetic Algorithm (GA) and Random Search (RS) were adopted, enhancing the quality of the radar signal. Based on radar measures of velocity and range, a machine learning approach was considered in order to classify objects detected by the radar. The Support Vector Machine (SVM) method was assumed to distinguish pedestrians and non-pedestrians objects. Four different SVM based models were considered in this work, aiming the improvement of both classification performance and speed, comparing three different kernel functions. As a result, it was possible to verify the advantages of the multi-objective optimization to extract the micro Doppler effects based on radar signals. Moreover, the optimization reached the velocity resolution of 0.12 m/s. To conclude, we showed that a model involving a polynomial kernel for the SVM reported better result in terms of accuracy (99.5%), confirming the promising perspectives of vehicles embedded radar-based safety systems.