I elaborate a bit on Toby Bartels answer:
Clearly, a smooth atlas gives a pointwise orientation via the coordinate vector fields, however, these may not be oriented (as local frames).
More precisely, for every $p \in M$ and (centered) coordinate chart $(U_p, \phi_p)$, we can choose the orientation of $T_p M$ to be the coordinate vector fields $$\left(\frac{\partial}{\partial \phi_p^1}{\big\vert_{p}}, ..., \frac{\partial}{\partial \phi_p^m}\big\vert_{p}\right\} =: \mathcal{U}_p,$$
where $\frac{\partial}{\partial \phi_p^i} := \phi^{-1}_{p\ast}(e_i) \in \Gamma(TU_p)$ and $e_i$ is the $i$th standard coordinate vector of $\mathbb{R}^m$.
Now, fix $p \in M$. Then, for every $q \in U_p$, the ordered set of vectors $(\frac{\partial}{\partial \phi_p^1}{\big\vert_{q}}, ..., \frac{\partial}{\partial \phi_p^m}\big\vert_{q})$ defines a basis for $T_q M$. However, one should note carefully that the orientation of $T_q M$ is determined by the tuple $\mathcal{U}_q$. So, it is meaningful to ask whether these two tuple belong to the same equivalence class, or in other words, agree in orientation. Whenever they do, then $\mathcal{U}_p$ is indeed an oriented local frame for $p \in M$, but, of course, it may happen that they do not agree. Similar holds for any other point of $M$.
So, saying that a pointwise orientation is continuous means essentially that the local orientations induced by $\{\mathcal{U}_{p}\}$ "patch together".
Addendum: It turns out that each coordinate frame constitutes an oriented local frame (with respect to the point-wise orientation induced by the coordinate charts) if and only if the transition functions are orientation-preserving.
In the abstract setting, then, we allow
- any choice of orientation for each $T_p M$, rather than the one induced by the atlas,
- any local frame, rather than the coordinate vector fields.
How one goes about showing with this definition that the Möbius band is non-orientable, I still do not know.