An Enhanced DarkNet53-based YOLOv3 Feature Pyramid Network for Real-Time Object Detection

With the rapid escalation of automatic and self-driving vehicles, the most efficient, safe, and reliable object detection system has been a big requirement for the automobile industry. These vehicles need an intelligent and highly accurate object detection and recognition system to be implemented for safe and secure driving to avoid road accidents. The ADAS is one of the features in automatic vehicles with improved navigation and object identification performance. Furthermore, better and optimal object detection is also required in surveillance and security along with industrial management and traffic management systems. The proposed work focuses on the development of an object detection model with enhanced and improved capabilities of the deep learning-based YOLOv3 model. This model has been supplied with the learning rate value of 0.001, training batch of 64, decay of 0.0005, and momentum has been set to 0.9. The activation function used in the downsample blocks has been set to either leaky ReLU or linear function. Further, the total number of classes on which the model has been trained is 80 including person, bicycle, car, parking meter, etc. This model is capable of object detection in the images collected from the web. This model has been trained with 80 different classes and testing has been done with the real-time data collected from the websites. This model is capable of dealing with object detection in those images varying from low to high density of the objects. The proposed model can be implemented in traffic light object detection, and ADAS along with security and surveillance applications for enhanced detection and optimal system generation.