Robust Detection of Small Moving Objects Against Real-World Complex Dynamic Natural Environments: Drosophila-Inspired...

Currently, small moving object detection remains a highly challenging problem, primarily attributable to four critical factors: limited pixel coverage, blurred texture features, indistinguishability from small-object-like background features (i.e., false positives), and vulnerability to environmental noise interference. The remarkable sensitivity of the Drosophila visual system to small moving objects, which originates from a specialized type of neuron known as "lobula columnar 11" (LC11), has provided inspiration for addressing this challenge. Current bio-inspired visual models have achieved certain advances. However, detection performance against real-world complex dynamic natural environments still requires further improvement. To address the challenge of limited detection accuracy for small moving objects against real-world complex dynamic natural environments, this paper proposes a Motion Small Object Detection (MSOD) model inspired by the Drosophila Vision Small Object Motion Sensitivity (DVSOMS) mechanism, namely DVSOMS-MSOD. The model consists of four stages: The first stage is preliminary processing of visual stimuli, where visual stimuli are perceived, converted to grayscale, and blurred. The second stage is the motion neural pathway, where visual signals are first decomposed into parallel ON and OFF neural pathway signals; then, the neural feedback mechanism is incorporated between the medulla and lobula complex, and the complete Hassenstein-Reichardt correlator (HRC) is integrated into the lobula complex; finally, the LC11 neuron is utilized to detect small moving objects and extract their location information. The third stage is the contrast neural pathway, where visual signals are first processed by the central and surrounding local neighborhoods, then local contrast information is calculated. The fourth stage is the integration of motion and contrast neural pathways, where the mushroom body generates motion trajectories using the location information of small moving objects, and subsequently generates contrast trajectories using the local contrast information and motion trajectories to more finely detect small moving objects. Under real-world complex dynamic natural environment datasets, compared with conventional machine learning methods for moving object detection, the proposed model achieved improvements of 77.82% and 78.70% in detection performance and output quality, respectively, while reducing running time by 10.60%. Compared with bio-inspired visual models for small moving object detection, the proposed model achieved improvements of 28.24% and 43.15% in detection accuracy and detection performance, respectively, but the running time increased by 43.40%. The proposed model demonstrates certain advantages in detection performance, output quality, and detection accuracy; however, its real-time performance still warrants further optimization.