Visual motion provides us with rich cues about the three-dimensional structure of our environment. However, it is generally unknown how circuits of neurons decode spatial information carried by patterns of visual motion. Here, we study the neural mechanisms of a collision avoidance behavior in walking Drosophila fruit flies as a simple model of motion-based spatial vision. In psychophysical experiments, we observed that flies exhibit slowing upon encountering small objects moving back-to-front in the frontolateral visual field. With a simple simulation, we demonstrate that this slowing can be seen as a behavior to avoid collisions with conspecifics that exploits the geometry of objects in near-collision courses. Next, we identified a visual neuron called lobula plate-lobula columnar type 1 (LPLC1) cells to be necessary and sufficient for the collision avoidance slowing behavior through synaptic silencing and optogenetic activation experiments. The visual response property of LPLC1 neurons measured with two-photon calcium imaging closely resembled the visual tuning of the collision avoidance behavior, notably in its spatially biased direction selectivity. Taking advantage of connectomic analyses, optogenetics, as well as neurochemical imaging and manipulations, we demonstrate that the peculiar visual tuning of LPLC1 is implemented through the pooling of elementary motion- and object-detecting neurons, as well as spatially biased glutamatergic inhibition. Additionally, we identified a downstream pathway of LPLC1 that mediate the collision avoidance behavior. Overall, our results exemplify how a small neural circuit can combine different visual features to solve a specific spatial vision problem, exploiting a universal geometrical constraint of the visual world.