Resource-Constrained Machine Learning for ADAS: A Systematic Review

Abstract

The advent of machine learning (ML) methods for the industry has opened new possibilities in the automotive domain, especially for Advanced Driver Assistance Systems (ADAS). These methods mainly focus on specific problems ranging from traffic sign and light recognition to pedestrian detection. In most cases, the computational resources and power budget found in ADAS systems are constrained while most machine learning methods are computationally intensive. The usual solution consists in adapting the ML models to comply with the memory and real-time (RT) requirements for inference. Some models are easily adapted to resource-constrained hardware, such as Support Vector Machines, while others, like Neural Networks, need more complex processes to fit into the desired hardware. The ADAS hardware (HW platforms) are diverse, from complex MPSoC CPUs down to classical MCUs, DPSs and application-specific FPGAs and ASICs or specific GPU platforms (such as the NVIDIA families Tegra or Jetson). Therefore, there is a tradeoff between the complexity of the ML model implemented and the selected platform that impacts the performance metrics: function results, energy consumption and speed (latency and throughput). In this paper, a survey in the form of systematic review is conducted to analyze the scope of the published research works that embed ML models into resource-constrained implementations for ADAS applications and what are the achievements regarding the ML performance, energy and speed trade-off. INDEX TERMS Machine learning, embedded software, automotive engineering, GPU, FPGA, ADAS. I. INTRODUCTION Safety is increasingly important for drivers and other users of the driving environment such as pedestrians, cyclists, bikers, scooters, etc. Only in 2016, road accidents in the European Union summed up to 25.600 fatalities and 1.4 million people injured [1]. Many measures are being taken by public institutions (improving road quality, enhancing driver consciousness and driving regulations, etc.) to improve these numbers. However, the cause of most of these accidents resides in human errors or distractions. For this reason, automotive firms are pushing up research on improving safety. The cornerstone of these advances are Advanced Driver Assistance Systems (ADAS). ADAS systems consist in several sensors, processing and actuators components that help drivers to avoid accidents and The associate editor coordinating the review of this manuscript and approving it for publication was Ye Duan.

A Glimpse of Physical Layer Decision Mechanisms: Facts, Challenges, and Remedies

“…All steps bring computational complexities along and require appropriate hardware capabilities. The systems may not meet high-level computational needs and constraint the solutions [57]- [60]. Therefore, the detection of system bottlenecks is critical for the selection of the proper machine learning algorithm.…”

Data Processing Approaches on SPAD-Based d-TOF LiDAR Systems: A Review

“…Gargoum et al divide possible features for pattern recognition in LiDAR data into on-road information, roadside information, and in conducting assessment of highway [64]. The property inference involved in ADAS can be categorized into 1) vehicle and pedestrian detection, 2) driver's state, behavior and identification, 3) traffic sign recognition, and 4) road detection and scene understanding [65]. A machine vision based traffic sign detection methods are reviewed in [66].…”

Accelerating deep neural networks for efficient scene understanding in automotive cyber-physical systems

“…In autonomous driving, Advanced Driver Assistance Systems (ADAS) rely on embedded systems with limited resources. ADAS is responsible of executing various machine learning tasks, including object detection, meaning that efficient implementations that take into account those limitations are critical [15]. To this end, singlestage detectors have been particularly studied for autonomous driving by either proposing specialized, compact deep models (e.g., [16], SqueezeDet [11], SA-YOLOv3 [17], Mini-YOLOv3 [18]) or applying MCA techniques [19] to existing, pre-trained models (e.g., [20] [21], [22], Efficient YOLO [23], ICME 2020 Competition [24]).…”