Counting dense object of multiple types based on feature enhancement.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Counting dense objects like pedestrians and vehicles in images is crucial but challenging.
• Existing methods primarily focus on single-class object counting in simpler scenes.
• Complex traffic scenes require simultaneous counting of multiple object types.

Purpose of the Study:

• To develop a new method for multi-class dense object counting in complex traffic scenes.
• To enhance feature extraction for improved classification and regression counting.
• To enable simultaneous counting of vehicles and pedestrians.

Main Methods:

• A novel multi-type dense object counting method based on feature enhancement.
• A counting model comprising a regression subnet and a classification subnet.
• The regression subnet generates two-channel density maps using enhanced features.
• The classification subnet assists by classifying dense vehicles and people.

Main Results:

• The proposed method successfully counts two types of dense objects simultaneously.
• It generates high-quality two-channel predicted density maps.
• Demonstrated superior counting performance compared to state-of-the-art methods on VisDrone+, ApolloScape+, and UAVDT+ datasets.

Conclusions:

• The feature enhancement approach effectively addresses multi-class dense object counting in complex scenes.
• The model achieves high accuracy in simultaneously counting vehicles and pedestrians.
• Future work will focus on expanding the model's capability to count a wider variety of objects.