Short-window crash prediction is a fundamental step in proactive traffic safety management that can monitor traffic conditions in real time, identify unsafe traffic dynamics, and implement suitable interventions for traffic conflicts. Short-window (e.g., hourly) traffic-collision count data, however, exhibits excessive zeros and serial autocorrelation. Most of the commonly used regression-based models fail to address excessive zeros and temporal structure simultaneously in hourly traffic-collision prediction. For example, Hurdle models and zero-inflated (ZI) models can address the overdispersion issue caused by excessive zeros but lack power to control the significant spatiotemporal characteristics inherent in the time-based collision data. To overcome these issues simultaneously, this paper develops a novel statistical model termed Zero-Inflated Logarithmic link for count Time series (ZILT) which is based on the framework of ZI models. Covariates (e.g., speed, vehicle type, and traffic volume) were extracted through deep-learning computer vision methods in vehicle detection and tracking on the image space. This new statistical model (i.e., ZILT) performs better at solving the issues of excessive zeros and serial dependencies. The prediction accuracy of the ZILT model improved by around 5% in relation to zero-inflated Poisson (ZIP) and Hurdle models. Results show that traffic crashes happening in the previous hour and other covariates such as truck-to-car ratio, holiday effect, traffic flow, and speed have significant influence on collision occurrence. Findings from this study could be utilized by relevant transport agencies in developing engineering interventions and countermeasures to proactively manage road safety.