Achieving fast and robust collision detection on autonomous agents requires perceiving threats and issuing early warnings with minimal latency. Particularly suited for this, dynamic vision sensors (DVS) capture high-speed motion with microsecond temporal resolution, empowering computational models of the lobula giant movement detector (LGMD) to perform rapid and selective collision detection. Aligning with this bio-inspired approach, the neuromorphic architecture and inherent temporal dynamics of spiking neural networks (SNNs) render them an excellent framework for integrating these advantages. However, existing systems still suffer from pronounced noise, highly variable event rates, and limited collision selectivity in complex motion scenarios. To overcome these limitations, we propose an SNN enhancing both robustness and selectivity. Inspired by retinal adaptation, we introduce an adaptive spatiotemporal filtering (ASTF) mechanism. Leveraging spatiotemporal integration and mixed-threshold neurons with global-local adaptation, the ASTF mechanism suppresses hot pixel and background activity (BA) noise and normalizes the output event rate, ensuring stable downstream processing. To further improve looming selectivity, we propose a depth-modulated spike-timing-dependent plasticity (D-STDP) learning rule. This mechanism incorporates a depth-modulated factor derived from global motion cues, which selectively potentiates synapses for looming motion, depresses them for receding motion, and gates off plasticity for irrelevant stimuli. We evaluate the proposed model on a multi-scenario dataset captured with a DVS, containing geometric, ball, and model vehicle motion. The results demonstrate that the proposed model achieves over 94% accuracy across diverse and challenging conditions, providing a promising approach for the development of bio-inspired collision detection visual systems.