Visual imagery is a critical perceptual source in human-robot interaction (HRI), extensively applied in various domains such as robotic navigation, target detection, and task planning. Aerial robots, also known as UAVs, are the typical representatives of teleoperation robots. The substantial distance between humans and remotely operated robots, coupled with the dynamic and open environments these robots navigate, exposes image-based HRI to significant security risks from both known and unknown threats. The trustworthiness of data sources is a fundamental prerequisite for ensuring the security of HRI systems. However, the multidimensionality of trust factors and the uncertainty of trust evidence have constrained its practical application. To advance this field, we have innovatively proposed a trust evaluation scheme based on multidimensional evidence fusion within the Belief Functions (BFs) framework. This scheme constructs an adaptive trust evaluation model capable of addressing both known and unknown threats. For known threats, we developed coarse-grained multidimensional trust elements and employed multiple lightweight SVM submodels to construct Basic Belief Assignments (BBAs), thereby achieving direct trust modeling and enabling rapid trust evaluation of images. Conversely, for unknown threats, we utilized pretrained models to establish fine-grained multidimensional trust elements, implementing indirect trust modeling through trust recommendation, thus facilitating indirect trust evaluation of images. We further combined direct and indirect trust to derive a overall trust assessment of the images, achieving adaptive dynamic trust evaluation in image-based HRI. Additionally, to more accurately represent image trustworthiness, we introduced a BBAs weighted fusion method, which aids in more rational trust aggregation. To validate the efficacy of our proposed method, we conducted experiments on a real aerial image dataset. The results demonstrate that our approach effectively characterizes the trustworthiness of visual imagery, thereby enhancing the security and reliability of HRI processes.