Otonom S r s i in ok-kategorili Kalabalik Analizi

With the ever-expanding interest in autonomous driving, the need for an accurate scene crowd analysis became essential. We exploit a CNN-based deep object counting and flow estimation method that utilizes density maps to estimate the distribution patterns of multiple target object classes, specifically vehicles, pedestrians, and bicycles that constitute key obstacles in driving. The CANnet2s deep network introduced for person heads is taken as the baseline architecture and it is adopted to multiple object classes by training from scratch. Video segments from the Waymo dataset, leveraging real-world urban frames captured under varying lighting and weather conditions are annotated and used for the training and inference. Performance evaluation results measured by MAE, RMSE and PSNR metrics demonstrate the network's capability to simultaneously process multi-category objects under diverse conditions including occlusion, pose and scale changes. Single object category evaluation performance is also reported for comparison.