Vision-Based Techniques for Identifying Emergency Vehicles

This paper discusses different computer vision techniques investigated by the authors for identifying Emergency Vehicles (EV). Two independent EV identification frameworks were investigated: (1) A one-stage framework where an object detection algorithm is trained on a custom dataset to detect EVs, (2) A two-stage framework where an object classification algorithm is implemented in series with an object detection pipeline to classify vehicles into EVs and non-EVs. A comparative study is conducted for different multi-spectral feature vectors of the image, against several classification models implemented in framework 2. Additionally, a user-defined feature vector is defined and its performance is compared against the other feature vectors. Classification outputs from each of the frameworks are compared to the ground truth, and results are quantitatively listed to conclude upon the ideal decision rule. As maintaining the speed of data output is the priority throughout our development, a computationally inexpensive object tracking algorithm is selected to accurately track EV between image frames. This vision-based EV detection scheme fused with data from other sensors on our autonomous vehicle shall be used to establish a sensor-fusion based EV detection and response framework in future work.