Multiperception enhanced network for efficient crowd counting

The task of crowd counting is to estimate the number of pedestrians in an image, and it is pivotal for applications in urban management and intelligent transportation systems. However, the performance of existing models is often hindered by significant challenges, namely scale variation and background interference. Moreover, deploying these models on resource-constrained edge devices for real-time analysis presents a substantial hurdle, necessitating a balance between accuracy and efficiency. To address these issues, we propose a lightweight multiperception enhanced network, termed multiperception enhanced network (MPENet). By contrast to existing computationally intensive approaches, MPENet presents a streamlined yet powerful framework designed for high performance. At its core, the network leverages a lightweight short for omni-scale network (OSNet) backbone for efficient feature extraction. To tackle the challenge of scale variation, we design a hierarchical scale perception module incorporating depthwise separable dilated convolutions. Concurrently, a focus enhancement module is introduced to mitigate background interference. This synergistic integration of modules facilitates the learning of robust, scale-aware features while maintaining a low computational overhead. Extensive experiments on widely used crowd and vehicle counting datasets validate that MPENet attains a superior balance between accuracy and efficiency. Consequently, this work provides a practical and effective solution for deploying high-performance crowd counting models on edge devices.