Detecting signals is often presented as a simple and solved problem. However, signal detection is typically not the only goal, as signal localization in time and frequency is often desired. This detection and localization process must also occur in the presence of non-ideal radio frequency (RF) receiver effects such as receiver non-flat noise floors, front-end inphase-quadrature (IQ) imbalance, local oscillator (LO) leakage, and in some cases power saturation. A robust signal detection system must overcome these hardware impairments and report only signals that correspond to true emitter devices. Additionally, some unspoken requirements take the form of detecting signals that are below the noise floor, very narrowband, very wideband, bursty, or frequency hopping. We present Searchlight, a signal detection system that solves these problems by using novel techniques that enable classical signal detection to work well using software defined radios (SDR). Performance results are presented for synthetically generated and over-the-air (OTA) data sets, which were obtained using commercially available hardware in a complicated signal scenario. Searchlight is shown to be an important enabler for true spectral monitoring platforms that desire to detect anomalous signals or apply machine learning (ML) algorithms to classify the various types of wireless activity in a given area.