To evaluate thermal imaging, we used a high-resolution thermal camera from Boson , with a 50° field of view (FoV) lens as a reference for validation and verification. As we do not know any specific emission values for nature in relation to solar radiation, we perform an automatic adjustment using histogram optimisation before each measurement interval. This allows us to achieve optimal contrast between the background and the deer in every scenario. Because these cameras are expensive, we also used HTPA-IR arrays from a Heimann Sensor . These are available in different resolutions, in our case 16 × 16 and 60 × 40 pixels. Since the detection of wild animals with thermal imaging cameras depends heavily on ambient light and temperature, the range was evaluated based on indoor tests, whereby range is understood to be the maximum distance at which a person can still be detected at a constant ambient temperature. The indoor comparison was considered appropriate because wildlife crossings mainly occur at twilight. These indoor tests therefore enable measurements to be taken without the influence of sunlight. Thermal solar radiation therefore has less influence. Standard Heimann sensors without adapted lenses were used for the tests. This means that the viewing ranges of the sensors differ (Table ).

For a first evaluation of detecting, a person was positioned in front of the sensors at the same distance and at different ambient temperatures. For example, as Fig.  shows, the person is easily recognisable at a distance of 14.5 m in a resolution of 60 × 40 pixels at an ambient temperature of 18 °C. However, at a resolution of 16 × 16 pixels, the person can no longer be identified due to the lower resolution in the same ambient temperature. This can be described by the spot size ratio (). If this is converted to the pixel area A, Eq. () is obtained. Here, d corresponds to the object distance andn corresponds to the number of pixels in the azimuth and elevation directions, respectively. However, using this equation to determine the maximum distance between objects in correlation to object size is incorrect. The reason for this is optical dispersion. This phenomenon describes how the thermal radiation, emitted from a small area, does not provide sufficient energy to the individual pixel. For a reliable measurement, the measuring point should therefore cover a resolution of at least 3 × 3 pixels . According to this theory, the target is already too far away for the sensor in Fig. a, because if the ambient temperature rises, only one pixel would show the person. Therefore, the further measurements are limited to the IR array with the 60 × 40 pixels. This sensor is therefore appropriate for the distance mentioned above, up to an ambient temperature of approximately 26 °C. 1A=d2⋅FoVazimuth⋅FoVelevation⋅π2nazimuth⋅nelevation⋅1802

Practical measurements show that the presence of deer at a reference ambient temperature of 20 °C can be recognised both in the IR array and in the high-resolution Boson camera (Fig. ). For our proof of concept, the standard configuration provided by Heimann, as shown in Table , is therefore sufficient. To extend the FoV to 90°, either a second sensor or a lens can be installed. However, as our evaluation shows, a larger FoV is linked to temperature resolution, which leads to lower distance resolution under the same environmental conditions. The single deer is 18 m away and, due to its body orientation, emits a large amount of heat, which the IR array maps over several pixels. By contrast, the deer on the left of the image is slightly farther away and aligned with the sensor system, resulting in a much lower heat signature.

As part of an initial trial, a convolutional neural network (CNN) was evaluated for its capacity to automatically detect deer that pose a threat to traffic safety. Due to the inability to clearly identify detection in radar data, the co-registration approach is to be utilised. Prior to the execution of the aforementioned procedure, it is necessary to first evaluate the thermal images. Given the simplicity and rapid detection rate of the YOLO (You Only Look Once) model, it was selected as the primary approach for the segmentation process. As in and , the YOLO CNN was used for automatic referencing. To achieve this, the 16-bit greyscale images from the Boson camera were cropped to 8 bits using mean-value nomination. This is necessary because thermal intensity is expressed as a relative number of 16 bits; thus, the measured relative temperature of each pixel is mapped to this value range. Additionally, YOLO is optimised for evaluating 8-bit RGB images. Due to the CNN's architecture, YOLO always requires three individual images (R, G, and B), which are superimposed to form the overall image. To analyse the impact of converting 16-bit thermal images to 8-bit false colours, YOLO was trained twice with the same images and data distribution (Table ). For this evaluation, a random selection of images was chosen, which were manually marked and provided with bounding boxes indicating where and how many animals were present. The selected images were then divided into two classes: with deer and without deer. Three sets of different sizes were formed from these classes (Table ). Due to the false-colour representation and the greyscale scaling, there are now six sets, three each for processing in the CNN.

After determining the range of the mean value and standard deviation, the single-channel greyscale image is used as a three-channel image in the first training case. This means the grey image is composed three times as RGB and transferred to YOLO (Fig. a). In the second training case, the greyscale image is coloured using a false-colour representation of the “jet” colour map and can be split into a three-channel RGB image (Fig. b) as a normal RGB image.

The recall is particularly noteworthy when assessing the training cases, as it demonstrates the number of deer that can be detected. The false-colour representation is 84.4%, about 3 points higher than the grey image. The precision evaluation shows an even smaller difference, with the false-colour representation at 85.4%, only 2 points higher. Deer are mainly not detected in the recall when they are at a great distance (20 m or more) from the measuring system or when a large part of the animal is obscured by vegetation.

This signifies that the acquisition of thermal data is especially advantageous in environments with substantial vegetation. In open terrain evaluating images in greyscale is sufficient to detect animals.

Radar detection depends strongly on carrier frequency, as higher frequencies reduce range but improve resolution. Free-space path loss (Eq. ) and range resolution (Eq. ) define sensor performance. For the FSPL, the maximum range r of interest is of primary importance, which depends directly on the carrier frequency fc, with c describing the speed of light as the propagation speed in air. 2FSPL=20log⁡10(r⋅fc⋅4πc) Due to the permitted bandwidths BW, a higher baseband frequency provides more accurate resolution ΔR . 3ΔR=c2⋅BW

The 60 GHz radar sensor from RFBeam with integrated signal processing enables a maximum range of r=30 m with a range resolution of ΔR=0.23 m and a velocity resolution of Δv=0.43 s (Fig. ) and still has an FoV of azimuth =elevation=±86°. This makes it suitable for installation in our sensor system. Using a moving target indication filter , static targets can be filtered without compromising the signal-to-noise ratio. This makes it possible to detect a deer at a distance of up to r=30 m.

In collaboration with tecVenture, we developed a custom 60 GHz radar sensor using the BGT60UTR11AIP chip . The system on chip (SoC) enables a simple system design. An STM32 serves as the micro-controller unit and configures the radar SoC. It also performs the initial FFT of the signal processing chain. The principle of a lens outlined from the IR arrays can be adapted here, and it has been implemented because the used chip has an FoV of 180° in azimuth and elevation. With the lens, the FoV can be limited to 70°. This results in a wider range of up to r=30 m. Due to the parameters used, a higher ΔR=0.075 m than with the RFBeam reference sensor can be achieved. In Fig. , three stationary deer were captured with the tecVenture sensor in a wildlife park. For this measurement, the MTI filter is deactivated, as the deer are located at static distances at feeding places.