Performance efficient vehicle detection and tracking based on pyramid pooling network: a review and implementation

The procurement of vehicles has grown over the years due to growth of the economy and improved living standards, resulting in environmental pollution and traffic jams. The objective is to detect individual vehicles with bounding boxes and also classify the type of vehicle. An edge computing model where processing is performed in the edge device itself is proposed. This paper presents a novel approach for vehicle detection and tracking using a pyramid pooling deep neural network. The proposed model initially extracts keyframes from input video to minimize processing overhead. These keyframes are then analyzed using the Spatial Pyramid Pooling (SPP-Net) detection algorithm to pinpoint object locations and classify vehicle types. It provides the coordinates, height and width of the bounding box in which the vehicle is present. It incorporates the appearance features and motion information, performing robust online tracking across consecutive frames and handling occlusions. Integration of spatial and temporal context enhancing consistency and smoothness in tracking. This approach optimizes object detection and tracking in video streams, enhancing efficiency and accuracy. Experimental results show a high accuracy improvement over RCNN and much faster training and inference making the proposed model a promising solution for real-world applications requiring accurate and reliable vehicle detection and tracking allowing inference directly on edge devices.