Road surveillance systems have an important role in monitoring roads and safeguarding their users. Many of these systems are based on video streams acquired from urban video surveillance infrastructures, from which it is possible to reconstruct the dynamics of accidents and detect other events. However, such systems may lack accuracy in adverse environmental settings: for instance, poor lighting, weather conditions, and occlusions can reduce the effectiveness of the automatic detection and consequently increase the rate of false or missed alarms. These issues can be mitigated by integrating such solutions with audio analysis modules, that can improve the ability to recognize distinctive events such as car crashes. For this purpose, in this work we propose a preliminary analysis of solutions based on Deep Learning techniques for the automatic identification of hazardous events through the analysis of audio spectrograms.