Detecting hot regions plays an important role in urban traffic planning and analytics, which is also useful in self-driving car routing and navigation. In this light, we propose and study a novel urban hot-region detection method by using massive geo-tagged image data. Given a set Q of regions (each region q is made up by a set of geo-tagged images), and a matching threshold θ , if a region q is matched with m other regions, its hot degree is defined by m. The hot-region detection (HRD) search finds the regions with the highest hot degrees. We believe that this type of search may benefit many applications in self-driving cars, including route planning and navigation, and traffic management and analytics in general. The HRD search is challenging due to two reasons. First, how to evaluate the similarity when matching different regions. Second, how to compute the HRD search efficiently, since its time complexity is O ( | Q | 2 ) . To overcome the challenges, we define a novel spatial-density correlation measure to evaluate the similarity between two regions, and develop a parallel search framework to process the HRD efficiently. In addition, a series of optimization techniques, e.g., pruning techniques, are defined to further enhance the query efficiency. Finally, we conduct extensive experiments on real data sets to study the performance of the developed methods.