EDCCN: A benchmark encoder-decoder framework for accurate crowd counting
Ankit Tomar, Rahul Nijhawan, Deepika Koundal · Neurocomputing · 2025
The increasing urban population has led to challenges in managing crowds in public places, especially in preventing tragic incidents like stampedes. Real-time accurate crowd counting (CC) in AI effectively manages crowd dynamics but faces significant obstacles such as background clutter, perspective variations, and occlusion. This study acknowledges the stated challenges in examining the effectiveness of convolutional arrangements by addressing the encoder–decoder crowd counter network (EDCCN). The model supports an integrated feature extraction process (segmented, edge-oriented, and texture), which makes it capable of calculating precise crowd counts in complex, dense situations. Its novel encoder–decoder arrangement explores low- and high-level crowd features in input images to address occlusion and uneven crowd distribution challenges in samples with different backgrounds. The EDCCN proposes two strategies to enhance people estimation accuracy: first, feature propagation guided without density maps to reduce background interference, and second, a complementary attention mechanism for improved information sharing among convolution layers . The EDCCN model harnesses intra- and inter-depth information representation through a non-increasing-order kernel arrangement, achieving state-of-the-art accuracy in people counting compared to existing methods across free (namely Mall, BRT, SmartCity, Indiana) and surveillance (JHU-Crowd, Venice, STech-B) datasets.