Object Detection and Tracking Approach for Traffic Monitoring

In present day, many object detection algorithms are available. These computer vision-based object detection algorithms help to detect, locate and trace an object from an image or a video. It requires high speed and accuracy along with efficiency when it is concerned with real-time systems. In machine learning and computer vision, object detection is considered to be an essential research area, also widely implemented in various sectors such as robotic navigation and intelligent video. The traditional approach to object identification consists of the steps as partitioning, clustering, feature extraction and classification. The solution is dependent of manual annotation, which leads to increase in the cost of the algorithm. Due to the diversity of objects, multiple models are needed for feature detection. As a result, classical object detection algorithms have poor generalizability, low detection accuracy, slow operating rate and low robustness. Object detection strategies are widly categorized as one-step and two-step object detection strategies. We are presenting a comparative analysis on object detection algorithms from two categories, i.e. single shot feed forward object detection algorithms and region proposal-based object detection algorithms. Under single shot feed forward, YOLOv7 is being is used along with pretrained weights trained on MS-COCO dataset from scratch. On the contrary, Mask R-CNN is being used to compare with, with its pretrained weights. This study presents a comparative analysis of YOLOv7 and Mask R-CNN in the context of accuracy, memory footprint and processing speed by retraining the models on the dataset of images and videos obtained through real-time systems in constrained.