Radar-based Vital Signs Detection for In-cabin Applications

Euro New Car Assessment Program (NCAP) is continually working to improve safety standards and is looking into innovative ways to enhance passenger safety. Euro NCAP and other NCAP standards are aiming to introduce incentives for in-cabin monitoring technology, such as CPD, which alerts guardians when a child is left alone in a car. They are also considering using in-cabin sensors to transmit occupants' vital signs to first responders after a crash [2].

CPD can be determined using the distinct characteristics of breathing rates. Therefore, determining vital signs plays an important role in implementing CPD.

Vital signs determination involves detecting breath rate (BR) and heart rate (HR). Breath rate plays a vital role in determining the presence of a child (CPD). In children, the breath rate ranges from 18 to 40 breaths per minute, and in adults, it's usually 12 to 16 breaths per minute [3]. Chest wall displacement due to breathing can range from a millimeter to a few millimeters. Heartbeat typically varies between 60 to 100 beats per minute and also affects the chest wall movement.

Radars play a crucial role in this application space, and this paper discusses a processing chain for in-cabin applications using an FMCW-based radar.

Figure 1 illustrates a typical example of a person standing in front of a radar. When the radar transmits a signal, it bounces off the person’s chest, and the radar receiver captures the reflected signal. By measuring the round-trip delay, we can estimate the distance of the target.

In addition, a displacement in the chest Δ𝑑 is observed, which encapsulates information about the breath and heart rate. The receiver captures information from the small chest displacement across multiple received signals. Using multiple observations of received signals, the information encapsulated in the small chest displacement is captured.

The transmitter sends multiple such signals/chirps, and the receiver gathers signals corresponding to each of these chirps. These received chirps are stacked on top of each other to form a 2D plane. A typical FMCW Single Input Multiple Output (SIMO) radar contains multiple receiver antennas, and the 2D planes formed at each receiver antenna are then combined to form a single 3D data cube.

The signal processing chain takes multiple 3D data cubes as input and processes them to extract information, such as the position and vital signs of the occupants. Applications with this functionality require a complex signal processing capability, which can be fulfilled by a high-performance digital signal processor (DSP). Using a programmable DSP offers various advantages, such as providing the desired adaptability and multi-capability support. The instruction set architecture (ISA) and memory subsystem of DSPs provide easy programmability and deliver sustained processing performance.

The Processing Chain section provides information about the signal processing required for an in-cabin application.

The processing chain is designed to extract the breath rate, heart rate, and estimated positions of occupants in the cabin, in addition to classifying the occupant as an adult or a child.

The input to the chain is a typical 3D data cube of the FMCW SIMO radar. The processing chain for vital signs differs from a typical external automotive radar as the displacement of the chest is in the order of sub-millimeter to millimeter range. Additionally, the slow movement of the chest implies that a longer observation duration is required than a traditional external automotive radar. This paper describes the signal processing chain designed considering typical FMCW radar parameters as follows:

A radar data cube is passed as input to the signal processing chain, as shown in Figure 2. The following sections describe the blocks (shown in Figure 2) in detail.

Clutter Removal

The clutter removal block helps remove static objects present inside the cabin, such as the seats, B-pillars, rear windshield, etc. These objects are stationary along the Doppler dimension and can be removed by applying a sample moving average (SMA) filter.

Range FFT

The clutter-removed time domain signal is then passed to the range FFT block. This block performs a Fast Fourier Transform (FFT) along the range bins to get the distance of the targets/occupants. The process involves using a windowing operation before the FFTs.

Range CFAR

The range profile is non-coherently integrated before it is passed to the CFAR module. A 1-D cell averaging constant false alarm rate (CA-CFAR) is performed to obtain the target positions and extract the valid range indices.

Angle FFT

This block performs FFT along the antenna direction of the range profile. The Angle FFT helps obtain the angle information and range-angle profile of the target.

Angle CFAR

Using the range-angle spectrum, a 1D CA-CFAR is performed along the angle direction to obtain the valid range and angle indices.

Pruning and Clustering

The pruning and clustering block takes the valid range-angle indices and the range-angle spectrum as input. It then compares the power of the indices with its neighbors from the range-angle spectrum to find the position of the local maximas. The local maximas that are close by are then grouped together to give the estimated range and angle positions of the target. These locations provide the positions of the occupants in the vehicle.

Vital Signs

Using each target's estimated distance and angle, a beam is constructed to extract all the Doppler samples at each target's estimated position. The phase of each Doppler sample is extracted, and an FFT is performed across the extracted phase. The peaks in the FFT spectrum represent the breath rate (BR) and the heart rate (HR). The BR and HR values provide the necessary vital signs. The determined BR is used to classify the occupant as a child for the CPD.

The In-cabin Use Case section discusses a typical in-cabin use case problem that could be solved using the above signal processing chain.

Figure 3 shows a typical in-cabin scenario. It represents a scenario where a radar is mounted on the front windshield. In this scenario, three adults and a child are present in the vehicle. In addition to picking up signatures from the occupants, the radar will also pick up signatures from unwanted cabin components, such as seats, B-pillar, rear windshield, and so on.

Figure 3 shows the occupants’ positions, seats, B-pillars, etc. Additionally, the occupants’ simulated breath and heart rates are shown.

The above-simulated scenario was provided as input to the processing chain. Figure 4 shows the results from the processing chain.

To perform the above processing for a CPD, Cadence^® Tensilica^® Vision 110 DSP was used. The following sections describe the Vision 110 architecture in brief and give an overview of the performance numbers of such an implementation.

The Cadence Tensilica Vision 110 DSP is based on an ultra-high-performance architecture designed for use in next-generation signal processing applications. Vision DSPs combine SIMD architecture with an up-to-five-issue VLIW processing pipeline and a rich and extensible set of interfaces. The Vision 110 DSP is built around a core vector pipeline consisting of up to one twenty eight 8-bit x 8-bit Multiply and Accumulate units (MACs), thirty-two 16-bit x 16-bit MACs, or optionally four or eight single-precision or 16 half-precision floating point or eight double-precision FMAs, along with a set of versatile pipelined execution units.

These units support flexible precision multiply-add, arithmetic and logical operations, data shift and normalization, data select, shuffle, and interleave. In addition to having the SIMD/VLIW DSP core, the Vision 110 DSP is built upon a 32-bit scalar processor, ideal for the efficient execution of control code. This combined SIMD/VLIW/Scalar design makes the Vision 110 DSP ideal for building real systems where high computational throughput combined with complex decision-making is required.

The Vision 110 DSP supports programming in C/C++ with a vectorizing compiler that allows automatic vectorization of scalar C code. It also supports vector gather and scatter operations to enable efficient vector processing of disparate data in local data memory. Gather and scatter operations read or write up to eight 16-bit or 8-bit vector elements or four 32-bit elements per cycle to arbitrary local data memory locations.

Tensilica DSPs also have a well-proven track record for handling such complex problems for various customers. For more information, refer to [4].

Performance is computed for occupancy sensing, and vital signs determination with the parameters mentioned in the Processing Chain section. The in-cabin use case scenario, as described, is used to measure the performance of the DSP and is summarized below.

The in-cabin use case problem is solved using the algorithm (described in the Processing Chain section), which takes approximately 600MHz (averaged over time) for ~50 frames per second (FPS) on a Cadence Tensilica Vision 110 DSP.

This document provides details of the signal processing chain that aids in-cabin sensing applications such as vital signs detection, occupancy detection. It also introduces the typical algorithms used. Performance results show that Vision 110 DSP with fixed-point and floating-point computations is best suited for in-cabin radar signal processing algorithms. A workspace containing optimized source code for in-cabin sensing implemented on Vision 110 will be made available upon request in the future.

For more information on the unique abilities and features of Cadence Tensilica IP and processors, refer to Compute IP.

References

[1] EURO NCAP RESCUE – TERTIARY SAFETY ASSESSMENT - ESV 2023

[2] Euro NCAP Vision 2030: A Safer Future for Mobility

[3] Understanding Normal Respiratory Rates in Children and Adults

[4] Vision 110 DSP User’s Guide