A Transfer-Learning-Based Approach for Emergency Vehicle Detection

Abstract

The paper presents a computer-vision based approach for real-time detection of different types of emergency vehicles under heavy traffic conditions. This enables preferential path clearance for emergency vehicles by the traffic controller which has the potential of saving lives, properties and increasing the ability to prevent crimes and drastically reducing the total time required by an emergency vehicle to reach its target destination. The main challenge emergency vehicles faced in and around the cities is heavy traffic jams, which significantly hampers their operations resulting in a disastrous outcome. In most of the cities, emergency vehicles are equipped with unique colors and sound system which enable the traffic police to identify them. As our cities become smarter and transition into an era of artificial intelligence, the old system may not be sustainable due to that fact that it needs humans to constantly monitor emergency vehicle arrival at the intersections and also the sound produces by such vehicles may be nuisance and discomforting to the general public. This paper proposed a method of automatic detection of four different categories of emergency vehicle irrespective of the vehicle’s shapes, models or manufacturers’ using modified version of YOLOv5 object detection algorithm. YOLO is an acronym for (You Only Look Once) and it is an object detection algorithm that divides images into a grid system. Each cell within the grid is responsible for detecting objects within itself. YOLO is one of the most famous object detection algorithms due to its speed and accuracy. YOLO models are used for object detection with high performance which consists of 84 classes to detect and differentiate between 84 different objects. The proposed model developed here is based on 4 classes which are (Firetrucks, Ambulance, Police Car, and Normal Cars) classes. The top layers (fully connected layers) of the YOLO algorithm were re-designed and retrained to get new learned weights while freezing the bottom layers (convolutional layers) and retaining the pre-trained YOLOv5 weights. After retraining with the proposed modified YOLOv5, the model has shown promising results and quite impressive metrics in detecting and classifying emergency vehicles and normal vehicles. Using Mean Average Precision (mAP) metric, for police cars we achieved 98%, 96% for fire trucks, 89% for ambulances and 97% for normal cars.

Keywords

• emergency vehicle
• yolov5
• deep learning
• object detection
• transfer-learning